From 327721b8eff19613901447a7c16e7ed3a42e467c Mon Sep 17 00:00:00 2001 From: api-gateway Date: Fri, 18 Sep 2026 16:33:33 +0800 Subject: [PATCH] =?UTF-8?q?=E7=AC=AC=E4=BA=8C=E5=8D=81=E4=B8=83=E8=BD=AE:?= =?UTF-8?q?=20=E4=B8=89=E7=BA=BF=E7=9B=B4=E4=BC=A0(/v1/messages,/v1/respon?= =?UTF-8?q?ses)+=E5=88=A0=E9=99=A4=E8=B7=AF=E7=94=B1=E5=88=AB=E5=90=8D?= =?UTF-8?q?=EF=BC=9B=E7=AC=AC=E4=BA=8C=E5=8D=81=E5=85=AB=E8=BD=AE:=20eleme?= =?UTF-8?q?nt-plus=20=E6=8C=89=E9=9C=80=E5=BC=95=E5=85=A5+=E6=AD=BB?= =?UTF-8?q?=E4=BB=A3=E7=A0=81=E6=B8=85=E7=90=86+=E5=89=8D=E7=AB=AF?= =?UTF-8?q?=E6=94=AF=E6=8C=81=E6=A0=BC=E5=BC=8F=E5=88=97/API=20Key=20?= =?UTF-8?q?=E6=96=87=E6=A1=A3?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- CHANGELOG.md | 73 ++ README.md | 126 ++- app/adapters/__init__.py | 11 +- app/adapters/openai_compat.py | 171 +++- app/config.py | 288 +++++-- app/main.py | 1076 ++++++++++++++++---------- app/metrics.py | 38 +- app/models.py | 60 +- app/settings.py | 90 ++- data/gateway.yaml | 200 +++-- docs/frontend-design.md | 4 + frontend/package-lock.json | 724 ++++++++++++++++- frontend/package.json | 3 +- frontend/src/lib/chat.ts | 3 +- frontend/src/locales/en.ts | 87 +-- frontend/src/locales/zh.ts | 87 +-- frontend/src/main.ts | 16 +- frontend/src/views/ApiKeys.vue | 27 +- frontend/src/views/Chat.vue | 56 +- frontend/src/views/Dashboard.vue | 34 +- frontend/src/views/GatewayConfig.vue | 3 +- frontend/src/views/Metrics.vue | 30 +- frontend/src/views/Models.vue | 133 ++-- frontend/src/views/Placeholder.vue | 20 - frontend/src/views/Settings.vue | 18 +- frontend/vite.config.ts | 25 +- live_check.py | 84 +- tests/test_gateway.py | 792 ++++++++++++------- 28 files changed, 3016 insertions(+), 1263 deletions(-) delete mode 100644 frontend/src/views/Placeholder.vue diff --git a/CHANGELOG.md b/CHANGELOG.md index 3400395..efcc676 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,78 @@ # CHANGELOG +## 2026-09-18 · 第二十七轮:三线直传 + **彻底删除路由别名** + +**老板决策**:wjd 路由这个功能去掉、删除掉,相关的代码和文档一起更新。 + +### 删除(路由别名 / fallback 链 —— 整块功能) +- `app/config.py`:删掉 `RouteInfo` 类、`GatewayConfig.routes` 字段、`find_route()` / + `all_routes()` / `validate_routes()` 三个方法,以及 YAML 里 `routes:` 段的解析。 + 老配置残留的 `routes` 段**不报错、直接忽略**(日志一条 warning),免得旧配置起不来。 +- `app/main.py`:删掉 `_route_chain()` / `_order_chain_by_speed()` / `_filter_chain_by_protocol()` / + `_vision_first()` / `_chain_has_vision()`;`_chat_fallback()` → `_chat_once()`(单模型 + 同模型 + 重试 + 熔断,不再"这个不行换下一个");流式 `_stream_response()` 单模型化(新增 `_sse_error()` + 统一错误帧);batch / 媒体端点 / 热重载 / 启动日志 / `/` / `/health` / `/v1/models` 里的路由痕迹 + 全部清掉。 +- `app/metrics.py`:删掉 `routes` 维度、`fallback_total` 与 `record_fallback()`;指标只剩 + **模型 / 供应商 / 账号** 三个维度。 +- `app/settings.py`:配置面板视图与写回不再有 `routes` 段。 +- `data/gateway.yaml`:`routes: {}` 整段删除,`server.retry` 注释去掉 route.retry 说法。 +- 限流从三级(全局/路由/供应商)变**两级(全局/供应商)**;`route.timeout`、路由级告警一并消失。 +- 前端:Models 页路由别名表 → 全部文本模型 + 一键测速;Chat 页去掉"路由"选项组与降级徽标; + Metrics 页「路由 vs 供应商」→「按供应商成功/失败」;Dashboard 路由链卡片 → 协议线覆盖卡片; + Settings 默认模型下拉去掉 `wjd`;ApiKeys 页示例改真实模型 id;中英文案同步(699 / 699 对称)。 + +### 保留 +- **一键测速**(Models 页每行的延迟实测)保留,改为按**模型自己的协议线**探测。 +- `wjd-` 开头的 API Key 前缀规则不变(那是 Key,不是路由)。 +- `app/routing.py` 保留 —— 它只做「model_id → 供应商 + 认证解析」,与被删的路由别名无关(名字沿用)。 + +### 为什么删(不是"藏") +三线直传落地后「入站格式 = 出站格式」,模型 id 必须与协议线一一对应;再套一层"自动挑最快模型" +的别名,会把调用方选定的**格式与能力**(视觉 / 工具 / 协议线)悄悄换掉。要换模型就换 `model`, +网关不替谁做主。 + +### 验证 +- 后端单测:**165 passed / 0 failed**(删 6 个路由引擎用例、改写 10 个、新增 2 个)。 +- 前端:`vue-tsc --noEmit` 0 错、`npm run build` 成功、`check-i18n` 中英 699/699 对称。 +- 生产**未动**:后端需 `supervisorctl restart api-gateway:api-gateway_00`(踢会话), + 前端 dist 覆盖需老板点头。 + +### 上线(老板批准重启,2026-09-18 15:16) +- 前端 `npm run build` 产物覆盖 `frontend/dist`(新入口 `index-CWUVm0px.js`,公网 + `https://api.wangjinding.com/web/...` 已服务新入口、旧入口 404)。 +- 后端 `supervisorctl restart api-gateway:api-gateway_00` → 新 pid,127.0.0.1:8808 正常, + `/health` 200、启动日志「已加载 3 供应商」。 +- **代码仍未提交 git**(老板明确要求"先别提交")。 + +### 上线后真机复验:抓到并修掉 3 个本轮新引入的缺陷(P0×3) +老板批准上线后,用 `live_check.py`(真实 key、走公网网关)复验,发现: +1. **【P0】三线直传入门即 500**:`_passthrough()` 里 `_enforce_rate_limit(cfg, request, None, pkey, model_in)` + 多传一个参数(函数只收 4 个)→ **每一次** `/v1/messages` 与 `/v1/responses` 都 `TypeError` → 500。 + 即本轮的核心功能两条线全灭(chat/completions 不受影响)。已改为 4 参调用。 +2. **【P0】`/metrics/prometheus` 500**:导出器仍读 `snapshot["fallback_total"]`(该字段已随路由/fallback 删除) + → `KeyError`。已删该指标行 + 其 HELP/TYPE,并顺带删掉 `/health` 里恒为 0 的 `fallback_total` 死字段。 +3. **【P0】图像/视频的 token 没发出去**:本轮改为"鉴权头按协议线在 `_req_headers` 组装"后, + `create_adapter` 不再往静态 headers 塞 `Authorization`,但媒体三个调用点仍在用 `self._headers` + ⇒ 图像生成/视频创建/视频轮询出站零鉴权头。真机:Agnes 图像 401「未提供令牌」,而同一 key + 直接 curl 同地址 **200**。已新增 `_media_headers()`(静态头 + Bearer + 会话头)并在三处使用。 +4. **测试基建(漏检根因之一)**:`tests/test_gateway.py` 的 `if __name__ == "__main__"` 运行块夹在文件中间, + 脚本模式下**只收集执行到该行时已存在的测试** → 定义在其后的用例根本不跑 + (`test_image_request_goes_to_requested_model` 白写了一轮)。已把运行块移到文件末尾, + 并新增 6 条**真发请求**的回归测试(`/v1/messages` 往返 + conversation_id 不透传、`/v1/responses` + + 跨线 400 闸门、直传供应商级限流 429、上游 4xx 原样回且不计熔断、流式字节原样转发 + 上游 4xx 真实状态码、 + Prometheus 200 且无已删指标、媒体三处带 Bearer)。 + 单测 **165 → 173 passed / 0 failed**。 +5. **文档订正**:README 残留的旧句「调用时 `model` 填 `wjd`,网关按顺序尝试」删除(路由已整块移除); + 鉴权小节「Key 均不落明文」改为与「Key 明文回显」取舍一致的表述。 + +### 复验结果(重启后再次真机) +`live_check.py`:**19 PASS / 0 FAIL / 0 WARN** —— 三线直传真实 200、跨线 400、`/metrics/prometheus` 200、 +图像生成 200。另手工验证:视频轮询可达上游(404「任务不存在」而非 401「未提供令牌」)、 +图像下载 `/v1/media/download` 560KB PNG 正常落盘、一键测速三条线各测一个模型均 ok +(chat/completions 7106ms、responses 2014ms、messages 1360ms)。 + + 按轮次记录 api-gateway 的修复与优化(README 侧重能力/配置说明,这里侧重每轮的改动、排查、验证与上线)。 --- diff --git a/README.md b/README.md index 55d01ae..1a85139 100644 --- a/README.md +++ b/README.md @@ -4,11 +4,12 @@ ## 特性 -- ✅ 多供应商统一路由(`model` 字段自动分发) -- ✅ 路由别名 `wjd`:运行时挑「当前可用 + 最近最快」的文本模型优先调用,其余按速度顺序 fallback + 熔断 -- ✅ 同模型重试 + 路由级超时 + **供应商熔断器**(chat 与媒体各自独立桶) +- ✅ 多供应商统一分发(`model` 决定供应商) +- ✅ **三线直传**:`/v1/chat/completions`(OpenAI)· `/v1/messages`(Anthropic)· `/v1/responses`(Responses)三条线各按**自家协议原样转发**,网关不翻译字段; + 模型支持哪几条线由 `models[].protocols` 声明,选错线**直接 400**并列出支持的线 +- ✅ 同模型重试 + **供应商熔断器**(chat 与媒体各自独立桶) - ✅ 图像 `/v1/images/generations` + 视频 `/v1/videos`(异步 + 等待模式 + Webhook) -- ✅ 页面账号体系鉴权(账号密码登录 + 每账号自建 API Key,三级限流**全局/路由/供应商**) +- ✅ 页面账号体系鉴权(账号密码登录 + 每账号自建 API Key,两级限流**全局/供应商**) - ✅ 用量指标 `/metrics` + **Prometheus** `/metrics/prometheus` + 访问日志 `/logs` - ✅ **Token 成本估算**(响应带 `estimated_cost_cny`) - ✅ **Batch 并发调用** `/v1/chat/completions/batch` @@ -23,9 +24,9 @@ api-gateway/ ├── data/gateway.yaml # 默认运行时配置(run.py 读取,hot_reload 开) ├── app/ -│ ├── config.py # YAML 加载 → dataclass(供应商/路由/鉴权/熔断/可观测/定价) +│ ├── config.py # YAML 加载 → dataclass(供应商/模型/协议线/鉴权/熔断/可观测/定价) │ ├── models.py # Pydantic 请求/响应模型(chat/image/video/batch) -│ ├── routing.py # model_id → 供应商路由 + 认证解析 +│ ├── routing.py # model_id → 供应商解析 + 认证解析(名字沿用;与「路由别名」无关) │ ├── metrics.py # 用量指标 + 访问日志 + 限流(无外部依赖) │ ├── circuit.py # 供应商熔断器(CLOSED/OPEN/HALF_OPEN) │ ├── main.py # FastAPI 主入口(全部端点 + 中间件) @@ -115,29 +116,43 @@ providers: 此外还有一层运行时兜底:上游若回 400 说某参数"只允许 X"(`invalid temperature: only 1 is allowed for this model`)或"必须大于 N",网关会自动按上游要求修正并**重发一次**,并把这个值记在适配器实例上(后续请求不再多跑一轮);日志会提示建议把它写进 `params`。管理员已写死的值不会被自动改掉(冲突时只记日志)。 -### 路由别名 + 按「可用 + 速度」动态排序 + 熔断 -(第十一轮 #4:`wjd` 不再按配置里的数组优先级,改为运行时挑可用且最快的文本模型先调用;详见下方路由说明) +### 三线直传(chat/completions · messages · responses)+ 熔断 + +三条协议线各有自己的入口,**入站格式 = 出站格式**:网关只做四件事 —— 鉴权、挑供应商、 +定点改 `model` 与强制参数、把响应(含 SSE)原样推回并读 usage 记账。**不做跨格式翻译**: +把 chat 线的模型打到 `/messages` 上,上游回的是 `500 Internal server error`;把 messages +线的模型打到 `/responses` 上回的是 `401 not supported for format` —— 两种都不像"格式选错", +所以网关**在入口直接 400** 并列出该模型实际支持哪些线,不让这种事拖成半天排查。 + +| 线 | 入站端点 | 出站鉴权 | +|---|---|---| +| `chat/completions` | `POST /v1/chat/completions` | `Authorization: Bearer` | +| `messages` | `POST /v1/messages` | `x-api-key` + `anthropic-version` | +| `responses` | `POST /v1/responses` | `Authorization: Bearer` | + +模型支持哪几条线写在各供应商的 `models[].protocols`;**不声明 = 只认 `chat/completions`** +(Agnes / 火山方舟即如此,行为与改造前完全一致)。OpenCode Go 的 29 个文档模型按**实测**逐一 +标线(`deepseek-v4.1-flash` / `deepseek-v4-pro` / `deepseek-v4-flash` / `deepseek-v4-flash-vision-exp` +三线全通)。`GET /v1/models` 会回 `protocols`,前端据此标格式。 + +**路由别名 `wjd` 已于第二十七轮(2026-09-18)整体删除**:不再有 `routes` 配置段、不再有 +多模型 fallback、`/v1/models` 里也没有 `type: "route"` 条目。理由:三线直传下模型 id 必须与 +协议线一一对应,再套一层"自动挑最快模型"的别名,会把调用方选定的格式与能力(视觉 / 工具 / +协议线)悄悄换掉。**要换模型就换 model id。** 老配置里残留的 `routes` 段会被忽略(日志一条 +warning),不会让进程起不来。限流因此变成**两级:供应商 > 全局**。 ```yaml -routes: - wjd: - timeout: 60 # 单模型 60s 没出结果就切下一个(避免 20×120s 挂死) - retry: 1 # 每个模型试 1 次,失败立即切下一个 - # wjd 链路 = 全部文本模型(新增/现有文本模型都必须加入) - models: - - "ark-code-latest" # 1) 火山方舟(最高优先级) - - "agnes-3.0-flash" # 2) Agnes 3.0 - - "agnes-2.5-flash" # 3) Agnes 2.5 - # ── 4-20) OpenCode Go 17 个文本模型(glm-5.3-flash…hy3, - # 完整列表见 data/gateway.yaml routes 段)── - circuit: # 供应商熔断 enabled: true failure_threshold: 3 # 窗口内失败 ≥3 次即熔断 cooldown: 30 # 熔断持续 30 秒 - window: 10 # 滑动窗口 = 最近 N 次调用(按次数,非按时间) + window: 10 # 滑动窗口 = 最近 N 次调用(按次数,非按时间) ``` +> **第二十七轮(2026-09-18):路由别名已整体删除** —— 下面这些路由相关修复 +> (`fallback_total`、`route.retry`、`routes` 维度指标、路由级限流/超时、路由告警) +> 随之一起移除,保留为历史记录。 +> > v2.3:`window` 现在真正按**最近 N 次调用**生效(之前是写死 60 秒、配置无效)。判定 = 最近 `window` 次里失败数 ≥ `failure_threshold`。 > > **修复(v2.5)**: @@ -209,12 +224,30 @@ circuit: # 供应商熔断 > - **无可达风险(已加约束说明)**:`release_probe()` 不校验令牌归属,理论上可能误还并发请求的探测令牌;逐一核对全部调用点后**当前不可达**(同 tick 或 except 分支内无抛错源),已在 docstring 写明"仅令牌持有者、且放行与归还间不得有 await"的前置条件,并给出后续放宽方案(probe token 归属校验)。 > - 审计同时确认**无新问题**:`json_or_raise()` 空/非 JSON body 映射、`Message.to_upstream()` 多轮 function calling、`stream` 严格校验、`unhandled_error_guard` 位于最外层且 500 带 CORS、`release_probe` 常规路径与 `probe_timeout` 自愈、`videos_poll` 404 归还不计失败、`_get_adapter` 移入 try 后的 503/404 映射、`_spawn_bg` 强引用与 lifespan 等待、`_opt_int` 的 `default>=minimum` 逐处正确。测试数 78 → 84(补 4 项回归:标准参数最终 payload、图像额外参数、流式 `in_flight` 时机、`base_url` 缓存 key)。 -调用时 `model` 填 `wjd`,网关按顺序尝试;某供应商连续失败会被熔断,期间快速跳过,让健康供应商顶上。 +> **修复(v2.11 · 第二十七轮上线后复验)**:老板批准重启上线后,用真机活体检查(`live_check.py` + 真实 key)复验, +> 抓到 **3 个本轮新引入的缺陷**(其中 2 个是"改了 A 处、漏了 B 处"的接线漏改),已全部修复并补测试: +> - **【P0】三线直传入门即 500**:`_passthrough()` 里调用 `_enforce_rate_limit()` 时多传了一个参数(5 参 vs 函数 4 参), +> 于是**每一次** `POST /v1/messages` / `POST /v1/responses` 都以 `TypeError` → 500「网关内部错误」结束 —— +> 正好是本轮的核心功能。真机表现为两条线全灭(chat/completions 线不受影响)。 +> - **【P0】`GET /metrics/prometheus` 500**:导出器还在读 `snapshot["fallback_total"]`,而该字段已随路由/fallback 链删除 +> → `KeyError`。Prometheus 抓取全断(`/metrics` JSON 不受影响)。已连同它的 HELP/TYPE 声明一起清掉, +> 并加断言:**导出的指标必须有 HELP 声明**(防"声明与导出脱节")。 +> - **【P0】图像 / 视频的 token 全部没发出去**:本轮把鉴权头改成"按协议线在 `_req_headers` 里组装"后, +> `create_adapter` 不再往静态 headers 里塞 `Authorization`,而媒体三个调用点(图像生成 / 视频创建 / 视频轮询) +> 仍在用 `self._headers` ⇒ 出站请求一个鉴权头都没有。真机表现:`POST /v1/images/generations` 上游 +> 401「未提供令牌」(同一个 key 直接 curl 同一个地址是 **200**)。已加 `_media_headers()` 统一补回。 +> - **测试基建(本轮漏检的根因之一)**:`tests/test_gateway.py` 的 `if __name__ == "__main__"` 运行块被夹在 +> **文件中间**,运行器只收集"执行到该行时已存在"的 globals ⇒ 定义在它之后的测试在 `make test`(脚本模式)下 +> 根本不会跑(`test_image_request_goes_to_requested_model` 白写了一轮)。已把运行块移到文件末尾, +> 并补齐 /v1/messages、/v1/responses 的**真发请求**测试(此前零覆盖)。测试数 **165 → 173**,全部通过。 +> - 复验结论:`live_check.py` **19 PASS / 0 FAIL / 0 WARN**(含三线直传真实 200、跨线 400、图像生成 200、 +> Prometheus 200);图像下载 `/v1/media/download` 实测 560KB PNG 落地正常。 ### 鉴权(页面账号体系 + 每账号 API Key) 鉴权**不写在配置文件里**:账号、密码、API Key 全部由管理台页面维护,持久化在 `data/users.json` -(密码 PBKDF2 哈希、Key SHA-256 哈希,均不落明文;文件权限 600 且已加入 `.gitignore`)。 +(密码只存 PBKDF2 哈希;Key 存 SHA-256 哈希用于鉴权,**另存一份明文**以便页面随时再复制 —— 取舍见下方 +「Key 明文回显」;文件权限 600 且已加入 `.gitignore`)。 `data/gateway.yaml` 只剩 `auth.session_ttl_minutes`(会话时长,默认 1440 分钟)一项。 ```yaml @@ -356,11 +389,13 @@ python run.py # 启动时自动读取 .env | `GET` | `/` | 否 | 端点导航 | | `GET` | `/health` | 否(可开) | 运维看板:uptime/失败率/耗时/熔断状态 | | `GET` | `/health/{provider}` | 否(可开) | 供应商探活 | -| `GET` | `/models` / `/v1/models` | 是 | 全部模型 + 路由别名 + 供应商 | -| `POST` | `/v1/chat/completions` | 是 | 主接口(fallback + 流式 + 熔断 + 三级限流 + 成本) | +| `GET` | `/models` / `/v1/models` | 是 | 全部模型(含视觉位与支持的协议线)+ 供应商 | +| `POST` | `/v1/chat/completions` | 是 | 主接口(chat/completions 线:流式 + 熔断 + 两级限流 + 成本) | +| `POST` | `/v1/messages` | 是 | Anthropic `messages` 线直传(同方协议原样进出) | +| `POST` | `/v1/responses` | 是 | OpenAI `responses` 线直传 | | `POST` | `/v1/chat/completions/batch` | 是 | 并发多 prompt | -| `POST` | `/v1/images/generations` | 是 | 图像生成(支持路由别名) | -| `POST` | `/v1/videos` | 是 | 视频创建(`wait`/`callback_url`,支持路由别名) | +| `POST` | `/v1/images/generations` | 是 | 图像生成(`model` 填真实模型 id) | +| `POST` | `/v1/videos` | 是 | 视频创建(`wait`/`callback_url`) | | `GET` | `/v1/videos/{id}` | 是 | 视频轮询(`?wait=true`) | | `GET` | `/metrics` | 是 | 用量统计(JSON,含熔断状态) | | `GET` | `/metrics/prometheus` | 是 | Prometheus 文本格式 | @@ -384,19 +419,19 @@ python run.py # 启动时自动读取 .env > 「仅会话」= 只接受账号密码登录换来的 `sess-` token,用 API Key 调用返回 403(见上文吊销语义)。 -> 媒体端点的 `model` 可以填**路由别名**(如 `wjd`):网关会自动解析到该路由链里第一个 `type==image/video` 的模型,找不到才 404。 +> 媒体端点的 `model` 必须是**真实模型 id**(第二十七轮起路由别名已删除,不再有别名解析)。 ## 关键行为说明(v2.4) -- **三级限流**:路由级 `routes..rate_limit_rps` > 供应商级 `providers..rate_limit_rps` > 全局 `server.rate_limit_rps`。路由/供应商级按独立桶分桶,打满则 429;chat 接口下打满会切下一个模型而不是直接 429。 +- **两级限流**:供应商级 `providers..rate_limit_rps` > 全局 `server.rate_limit_rps`。供应商级按独立桶分桶,打满则 429(chat / 媒体共用 `prov:` 桶);路由级限流随路由功能删除。 - **媒体端点也限流**(v2.3 / v2.6 修正):`/v1/images`、`/v1/videos` 按 `providers..rate_limit_rps` 的**供应商桶**限流,与 chat 共用同一个 `prov:` 桶(v2.6 起媒体不再另起 `media:prov:` 独立桶 —— 那样实际放行量会翻倍、与文档不符)。高成本资源不再裸奔。 -- **route.timeout 全链生效**(v2.3):配了 `routes..timeout` 后,**整条链(含链首最高优先级模型)统一用这个超时**;之前只对兜底模型生效、链首仍用 `server.timeout`。 +- **协议线闸门**(第二十七轮):`POST /v1/messages`、`/v1/responses` 只放行声明了对应线的模型,否则 400 并列出该模型支持的线;网关不做跨格式翻译。 - **熔断按次数窗口**(v2.3):`circuit.window` 现在真正按"最近 N 次调用"滑动(之前写死 60 秒、配置无效)。判定 = 最近 window 次里失败 ≥ failure_threshold。 - **失败也算耗时**:`/metrics` 的 P50/P95 包含失败请求的等待时间。 - **in_flight 精确统计**(v2.4):按响应**实际类型**(`StreamingResponse`)判断流式,非流式/413/502 在中间件减计数,SSE 流式在生成器收尾减计数。修复了"3MB 的流式 POST 被 413 后 in_flight 永久 +1"的泄漏。 - **适配器 client 并发安全**(v2.4):`get_client()` 加 `asyncio.Lock`,并发下只建一个 client(之前竞态建多个泄漏 fd);`close()` 幂等可重入(之前并发 double-close 抛 `Event loop is closed`)。 - **图像 size 缺省 1K**(v2.4):`/v1/images/generations` 不传 size 时默认最低档 `1K`,避免 Agnes 因缺 size 返回 400;`n=1` 时省略该字段(Agnes 一次一张)。 -- **热重载真正生效**:改 `circuit` / `observability` / `metrics` / `auth` / `routes` / `pricing` 即时重建运行时对象;日志级别/JSON 格式热改也真能生效;`uptime_seconds` 因热重载不会清零。 +- **热重载真正生效**:改 `circuit` / `observability` / `metrics` / `auth` / `pricing` / 模型与协议线即时重建运行时对象;日志级别/JSON 格式热改也真能生效;`uptime_seconds` 因热重载不会清零。 - **优雅关闭**:`server.graceful_shutdown` 传给 uvicorn `timeout_graceful_shutdown`,收到停止信号时先排空在途请求再退出。**(v2.8 重要修复)**:lifespan 里曾用 `loop.add_signal_handler` 注册自己的 SIGTERM/SIGINT 处理器,把 **uvicorn 自己的处理器覆盖掉了**,导致 `SIGTERM`/`SIGINT` 完全无法停服(实测发信号后 10s+ 仍存活,只能 `kill -9`)。现已移除该注册,交给 uvicorn 处理;实测 SIGTERM 后 0.5s 内完成 `Shutting down → 应用收尾 → Finished`。 - **OpenAI 全兼容**(v2.2):响应带 `created` / `system_fingerprint`,流式末尾带 `usage` 块,错误为 `{"error":{message,type,code}}`,`/v1/models` 标准端点——OpenAI 官方 SDK 零适配直连。 @@ -411,10 +446,15 @@ curl -X POST http://localhost:8808/v1/chat/completions \ -H "$AUTH" -H "Content-Type: application/json" \ -d '{"model":"agnes-3.0-flash","messages":[{"role":"user","content":"你好"}],"max_tokens":256}' -# 路由别名 wjd(运行时挑可用 + 最快文本模型优先,其余按速度顺序 fallback + 熔断) +# 换模型 = 换 model id(第二十七轮起没有路由别名,网关不会替你挑模型) curl -X POST http://localhost:8808/v1/chat/completions \ -H "$AUTH" -H "Content-Type: application/json" \ - -d '{"model":"wjd","messages":[{"role":"user","content":"hi"}]}' + -d '{"model":"deepseek-v4.1-flash","messages":[{"role":"user","content":"hi"}]}' + +# Anthropic messages 线(同方协议直传,鉴权头换成 x-api-key) +curl -X POST http://localhost:8808/v1/messages \ + -H "$AUTH" -H "Content-Type: application/json" \ + -d '{"model":"deepseek-v4.1-flash","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}' # 并发多 prompt(batch) curl -X POST http://localhost:8808/v1/chat/completions/batch \ @@ -451,10 +491,10 @@ curl "http://localhost:8808/v1/videos/?model=agnes-video-2.5-flash&wai - 里程碑:**M1** 登录 + 网关托管 + 去掉批量页(构建栈 Vue 3 + Vite + TS + Element Plus + ECharts + Pinia + Vue Router + marked)。JS 主包约 1.2MB,分包/按需引入留给 M5。 - 里程碑:**M2** 对话 Playground(`/web/#/chat`):流式/非流式对话(fetch 直连 - `/v1/chat/completions` 解析 SSE)、模型/路由选择(路由显示完整降级链)、参数面板 + `/v1/chat/completions` 解析 SSE)、模型选择(第二十七轮起是模型 + 协议线,不再有路由别名)、参数面板 (temperature / top_p / max_tokens / 流式开关 / 系统提示词)、Markdown+代码高亮渲染 (marked + highlight.js + DOMPurify 消毒)、推理折叠(`reasoning_content` 单独折叠)、 - 降级可视化(meta 帧的 served/provider/fallback_used 徽标)、token/成本展示(流式 usage + 服务信息徽标(meta 帧的 served/provider;`fallback_used` 随路由功能删除)、token/成本展示(流式 usage 帧现也带 `estimated_cost_cny`)、会话历史 sessionStorage 持久、Enter 发送/Shift+Enter 换行。 - 里程碑:**M3** 图像页面(`/web/#/images`)+ 视频页面(`/web/#/videos`): 图像生成网格(`/v1/images/generations`,url/base64 均可显示、el-image 大图预览、耗时/供应商标注、 @@ -467,7 +507,7 @@ curl "http://localhost:8808/v1/videos/?model=agnes-video-2.5-flash&wai - 里程碑:**M4** 仪表盘图表打磨 + 指标页 + 日志页(引入 ECharts,`src/components/EChart.vue` 封装 + `src/lib/echarts.ts` tree-shake 按需注册): 仪表盘新增供应商份额饼图 + 模型延迟 P50/P95 柱图;指标页(`/web/#/metrics`)含指标卡 - (已运行/在途/降级/双 token/调用比)+ 各模型 ok/fail 堆叠柱图 + 供应商份额 + 路由 vs 供应商 + (已运行/在途/双 token/调用比)+ 各模型 ok/fail 堆叠柱图 + 供应商份额 + 按供应商成功/失败 堆叠 + 延迟 P50/P95 + 熔断器状态表(自动刷新 8s);日志页(`/web/#/logs`)展示最近 200 条 `/logs` 访问日志(方法/路径/供应商/实际模型/OK-FAIL/SSE/调用方/rid/错误),支持方法、结果、 关键字筛选与自动刷新 5s。EChart 共享分包约 594kB。 @@ -628,7 +668,7 @@ curl "http://localhost:8808/v1/videos/?model=agnes-video-2.5-flash&wai - **图表文字挤压修复**:ECharts 6 把 legend 默认位置从 top 改为 bottom,导致图例压在 x 轴标签上。新增 `src/lib/charts.ts` 统一 `legend(top:0)` / 类目轴截断+倾斜+hideOverlap / `grid.containLabel` 留白,仪表盘与指标页 4 类图全部接入。 -- **模型与供应商页**(`/web/#/models`):指标卡(模型/文本/图像/视频/路由/供应商)+ 路由别名表 +- **模型与供应商页**(`/web/#/models`):指标卡(模型/文本/图像/视频/供应商)+ 模型清单(ID/类型/供应商/视觉/延迟,一键测速按模型自己的协议线探测),数据来自 `/v1/models` + `/health`。(第二十七轮:路由别名表已随功能删除) (降级链逐 tag 展示、供应商数)+ 模型清单(ID/类型/供应商/Key 已配),数据来自 `/v1/models` + `/health`。 - **多语言(M5+)**:Vue I18n + `src/locales/zh|en.ts`,全站中英可切换;语言按钮在头部与登录页、设置页 (`settings.language`),存 `localStorage.gw_locale`、默认中文;Element Plus 语言包经 @@ -745,7 +785,7 @@ curl "http://localhost:8808/v1/videos/?model=agnes-video-2.5-flash&wai > - **清空/清除类操作补二次确认**:视频页「清除已完成」、图像页「清空历史」都是**服务端真删且不可逆** > (不是只清本地视图),原先点一下就没了。现在弹确认框并写明条数与后果(另:视频页没东西可清时按钮置灰)。 > - **设置 → 关于去掉文档外链**(不再往外部站点跳),同时删掉对应 i18n 键。 -> - **API 密钥页新增「API 调用说明」**:接口地址(复制按钮)· 鉴权头 · 模型选择(`wjd` 路由别名 vs 真实模型 id)· +> - **API 密钥页新增「API 调用说明」**:接口地址(复制按钮)· 鉴权头 · 模型选择(填真实模型 id;路由别名已于第二十七轮删除)· > 请求格式 · 常用端点表(与 `app/main.py` 的真实路由一一对应)· 四段可直接复制的示例 > (对话 / 图像 / 视频提交+轮询 / Python OpenAI SDK)· 七条注意事项(n=1、seconds 4–12、 > 轮询别快于 5s、素材可用 data URI 但请求体 <2MB、Key 调管理接口 403、Key 泄漏处置、审计归属)。 @@ -782,8 +822,8 @@ curl "http://localhost:8808/v1/videos/?model=agnes-video-2.5-flash&wai > 熔断桶与调用指标都不动,前端按退避继续轮(页面轮询间隔保持 5s,不按文档建议的 1~2s,正是为了不撞限流)。 > - **视觉能力位**:`/v1/models` 新增 `vision`(`agnes-3.0-flash` = true,即文档里的"输入模态:文本、图像 URL"; > OpenCode Go 侧 `deepseek-v4.1-flash` 与 `deepseek-v4-flash-vision-exp` 实测同为 true); -> **带图请求走路由别名 `wjd` 时把视觉模型排到链首**(否则链首的纯文本模型要么忽略图片、要么报参数错误, -> 两种都极难自查);直接指定真实模型时不改用户选择。 +> **带图请求不再重排模型**(第二十七轮:路由链已整体删除,网关不替用户换模型)——选了不支持视觉 +> 的模型就照原样发上去,前端提前给告警条并提供一键切换,网关侧只记一条日志。 > - 回归测试新增 6 组(vision 暴露 / 视觉优先重排 / 带图块识别 / 图像 n 拦截 / 图像 4xx·5xx 分流 / > 视频 videos·size·aspect·n 校验 / 轮询 429),单测 **124 passed / 0 failed**、pyflakes 0; > 前端 `npm run typecheck`(vue-tsc)**0 error** —— 顺手补了 element-plus 语言包的模块声明、 @@ -1191,7 +1231,7 @@ curl "http://localhost:8808/v1/videos/?model=agnes-video-2.5-flash&wai ## 测试 ```bash -# 单元测试(mock 供应商,不依赖真实网络,覆盖配置/路由/熔断/适配器/端点) +# 单元测试(mock 供应商,不依赖真实网络,覆盖配置/协议线/熔断/适配器/端点) python tests/test_gateway.py # 或 pytest tests/ -q @@ -1203,12 +1243,12 @@ pytest tests/ -q # 冒烟:健康 / 模型列表 / chat / batch / 错误格式(网关未起或无 key 时优雅 SKIP) python smoke_test.py --url http://127.0.0.1:8808 -# 活体检查:真实上游(Agnes / 火山方舟)跑 14 项端到端行为校验,含路由 fallback、 +# 活体检查:真实上游(Agnes / 火山方舟 / OpenCode Go)端到端行为校验,含三线直传与跨线 400、 # 流式 SSE、413、CORS 预检、图像生成等 python live_check.py --url http://127.0.0.1:8808 ``` -> 当前状态:单元测试 **168 passed / 0 failed**(`python tests/test_gateway.py`,不需要 pytest);冒烟 5/5 PASS;活体检查 14 PASS / 0 FAIL / 0 WARN。 +> 当前状态:单元测试 **165 passed / 0 failed**(`python tests/test_gateway.py`,不需要 pytest);冒烟 5/5 PASS;活体检查 14 PASS / 0 FAIL / 0 WARN。 ## 备注 diff --git a/app/adapters/__init__.py b/app/adapters/__init__.py index b5d55b1..01a038d 100644 --- a/app/adapters/__init__.py +++ b/app/adapters/__init__.py @@ -12,13 +12,17 @@ def create_adapter( ) -> BaseAdapter: """根据供应商配置创建对应适配器。 - 当前三个供应商(Agnes / 火山方舟 / OpenCode Go)均为 OpenAI 兼容协议, + 当前四个供应商(Agnes / 火山方舟 / OpenCode Go)均为 OpenAI 兼容协议, 统一使用 OpenAICompatAdapter,通过 base_url 和 extra_headers 区分。 图像/视频端点通过 media 配置覆盖(Agnes 走 /videos + /agnesapi)。 OpenCode Go 由 requires_session_header 触发每请求动态注入 x-opencode-session。 + + 第二十七轮(三线直传): 鉴权头不再在这里拼死,而是把裸 token + 供应商的 + `endpoints`(每条线的路径/鉴权方式/固定头)交给适配器,由它在**每次请求**时按 + 入站协议线组装(OpenCode Go 的 messages 线认 x-api-key,另两条线认 Bearer)。 + 这样同一个供应商的三条线共用一个适配器与连接池,也不需要在缓存 key 里区分协议。 """ headers = { - "Authorization": f"Bearer {token}", "Content-Type": "application/json", } headers.update(pconfig.extra_headers) @@ -36,4 +40,7 @@ def create_adapter( # 不会出现"主链路修了、测速还漏着"的偏差。 # 配置变更时 main._apply_config 会丢弃整个适配器缓存,约束不会滞留旧值。 model_params={m.id: dict(m.params) for m in pconfig.models if m.params}, + # 第二十七轮: 三条线的出站地址与鉴权方式(直传与常规调用共用) + token=token, + endpoints=pconfig.endpoints, ) diff --git a/app/adapters/openai_compat.py b/app/adapters/openai_compat.py index ea490cb..260bdcd 100644 --- a/app/adapters/openai_compat.py +++ b/app/adapters/openai_compat.py @@ -14,12 +14,18 @@ import json import logging import re import uuid +from contextlib import asynccontextmanager from typing import Any, AsyncIterator import httpx from app.adapters.base import BaseAdapter, json_or_raise -from app.config import MODEL_PARAM_KEYS +from app.config import ( + PROTOCOL_CHAT, + EndpointSpec, + MODEL_PARAM_KEYS, + _default_endpoints, +) from app.models import CHAT_CORE_KEYS, PROTECTED_VIDEO_KEYS, strip_protected log = logging.getLogger("api-gateway") @@ -38,7 +44,13 @@ _PARAM_MIN_RE = re.compile( class OpenAICompatAdapter(BaseAdapter): - """适用于所有 OpenAI 兼容 /chat/completions 端点的适配器。""" + """适用于所有 OpenAI 兼容 /chat/completions 端点的适配器。 + + 第二十七轮(三线直传)后它同时承担**协议直传**职责:同一家供应商的三条线 + (chat/completions、messages、responses)共用这一个适配器、共用一个连接池, + 差别只在「路径 + 鉴权头」——由 provider 的 `endpoints` 配置提供(见 config.EndpointSpec)。 + 直传路径**不翻译任何内容**:入站 body 原样出去,上游响应原样回来。 + """ def __init__( self, @@ -49,9 +61,17 @@ class OpenAICompatAdapter(BaseAdapter): media=None, requires_session_header: bool = False, model_params: dict[str, dict[str, Any]] | None = None, + token: str = "", + endpoints: dict[str, EndpointSpec] | None = None, ): super().__init__(name=name, base_url=base_url, timeout=timeout, media=media) self._headers: dict[str, str] = headers or {} + # 第二十七轮: 每条协议线的出站配置(路径 / 鉴权方式 / 该线固定头)。 + # 不传 = 三条标准线的默认值,所以旧调用点(含单测直接 new)行为不变。 + self._endpoints: dict[str, EndpointSpec] = dict(endpoints or _default_endpoints()) + # 裸 token:每条线要发的鉴权头不同(Bearer / x-api-key),所以不能在构造时就把 + # header 拼死。没显式传就回头从静态 headers 里解析(兼容直接 new 的调用方)。 + self._token = token or self._token_from_headers(self._headers) # OpenCode Go 要求每个 /chat/completions 带稳定 x-opencode-session。 # 不能放静态 headers(_get_adapter 会缓存适配器、冻结 header), # 所以在每次请求时动态生成一个 UUID 会话头。 @@ -64,6 +84,45 @@ class OpenAICompatAdapter(BaseAdapter): # 仅作兜底 —— 管理员在 gateway.yaml 写死 params 后就不再需要它。 self._learned_params: dict[str, dict[str, Any]] = {} + @staticmethod + def _token_from_headers(headers: dict[str, str]) -> str: + """从静态 headers 里反解 token(兼容历史调用方式:构造时就带 Authorization)。""" + auth = headers.get("Authorization") or headers.get("authorization") or "" + if auth.lower().startswith("bearer "): + return auth[7:].strip() + return headers.get("x-api-key") or "" + + # ── 协议线(第二十七轮) ───────────────────────── + + def endpoint(self, protocol: str) -> EndpointSpec | None: + """取某条线的出站配置(None = 该供应商没有这条线)。""" + return self._endpoints.get(protocol) + + def _path(self, protocol: str) -> str: + ep = self._endpoints.get(protocol) + if ep is None: + # 走到这里说明配置校验漏了(validate_config 会拦「模型声明了但供应商没配」), + # 抛清晰错误而不是退化成默认路径 —— 猜出来的地址只会收到 404,更难查。 + raise ValueError(f"供应商 {self.name} 未配置协议线 {protocol}") + return ep.path + + def auth_headers(self, protocol: str) -> dict[str, str]: + """按该线的鉴权方式组装鉴权头(Bearer / x-api-key / 无)。 + + 这是三线直传里"唯一必须做的 glue":同一个供应商的三条线鉴权方式可能不同 + (实测 OpenCode Go:chat/completions 与 responses 认 `Authorization: Bearer`, + messages 认 `x-api-key`;用 Bearer 打 /messages 会 401 Missing API key)。 + """ + ep = self._endpoints.get(protocol) + style = (ep.auth if ep else "bearer") + out: dict[str, str] = {} + if style == "bearer" and self._token: + out["Authorization"] = f"Bearer {self._token}" + elif style == "x-api-key" and self._token: + out["x-api-key"] = self._token + if ep: + out.update(ep.headers or {}) # 如 anthropic-version + return out # ── 模型级强制参数 / 上游参数约束自学习(第二十五轮) ── def _forced_params(self, model: str) -> dict[str, Any]: @@ -128,19 +187,103 @@ class OpenAICompatAdapter(BaseAdapter): model, changed) return self._forced_params(model) - def _req_headers(self, session_id: str | None = None) -> dict[str, str]: - """每次请求的 headers。需要会话头的供应商动态注入 x-opencode-session。 + def _req_headers(self, session_id: str | None = None, + protocol: str = PROTOCOL_CHAT) -> dict[str, str]: + """每次请求的 headers:静态头 + 本条线的鉴权头 + 会话头。 session_id 由调用方传入(网关把对话的 conversation_id 透传过来): - - 有 → 同一段对话的所有请求共用同一个 session(吃到上游提示词缓存); + - 有 → 同一段对话的所有请求共用同一个 session(上游据此做路由与缓存); - 无 → 回退为每请求一个随机 UUID(仍能过上游的强校验,只是没有缓存收益)。 + + 第二十七轮实测:OpenCode Go 的**三条线都硬性要求该头**(缺任一条 → 400 + MissingSessionID),所以注入逻辑与协议线无关,统一在这里做。 """ - if not self._requires_session: - return self._headers headers = dict(self._headers) - headers["x-opencode-session"] = session_id or uuid.uuid4().hex + # 鉴权头按线走:先清掉静态头里可能残留的鉴权(历史写法:create_adapter 曾把 + # Authorization 塞进静态 headers),再由 auth_headers 按该线的方式补上。 + headers.pop("Authorization", None) + headers.pop("authorization", None) + headers.pop("x-api-key", None) + headers.update(self.auth_headers(protocol)) + if self._requires_session: + headers["x-opencode-session"] = session_id or uuid.uuid4().hex return headers + def _media_headers(self) -> dict[str, str]: + """媒体端点(图像生成 / 视频创建 / 视频轮询)的 headers。 + + 媒体端点不属于「三条协议线」,所以不走 `_req_headers` —— 历史口径是固定用 Bearer。 + 第二十七轮把鉴权头改成「按协议线在 `_req_headers` 里组装」之后,这三个调用点却还在 + 用静态 `self._headers`,而 `create_adapter` 已经不再往静态头里塞 Authorization + ⇒ **token 一个都没带出去**。真机表现:`POST /v1/images/generations` 上游 401 + 「未提供令牌」(同一个 key 直接 curl 同一个地址是 200)。 + 这里统一补回来:静态头 + Bearer(有 token 时)+ 会话头(供应商要求时)。 + """ + headers = dict(self._headers) + if self._token: + headers["Authorization"] = f"Bearer {self._token}" + if self._requires_session: + headers["x-opencode-session"] = uuid.uuid4().hex + return headers + + # ── 协议直传(第二十七轮)──────────────────────── + # 「入站是什么格式,出站就是什么格式」:这两个原语把入站 body **原样**发到该线的上游 + # 地址,**不翻译任何字段**。网关只做三件事:① 挑对供应商与线;② 把 model 改成真实 + # id 并盖上模型级强制参数;③ 原样把响应(含 SSE)交回客户端。 + # 中间不做 has_images 之类的语义判断,也不解析 body 结构。 + + def prepare_payload(self, model: str, payload: dict[str, Any]) -> dict[str, Any]: + """直传前的定点改写:model 换成真实 id + 盖模型级强制参数(其余一律不动)。 + + - model:入站就是真实模型 id(第二十七轮起路由别名已删除,不再有别名解析, + 网关也不会替你换模型)—— 这里仍保留"定点改写 model"的能力:调用方写的 id + 必须落到某条线能认的 id 上; + - 强制参数:`models[].params`(如 kimi-k2.7-code 只接受 temperature=1)在直传里 + 同样必须生效,否则"直传"就成了"漏掉网关的硬约束"。 + 这两处是**唯一**允许改写入站 body 的地方(改的都是"发给谁/必须用什么参数", + 不是协议语义),其余字段一律原样。 + """ + out = dict(payload) + out["model"] = model + forced = self._forced_params(model) + if forced: + out.update(forced) + return out + + async def raw_json(self, protocol: str, payload: dict[str, Any], + session_id: str | None = None) -> tuple[int, Any]: + """非流式直传:返回 (上游状态码, 上游 JSON/文本)。 + + 刻意**不 raise_for_status**:直传口径下上游的错误体要原样交回客户端 + (老板决策:"其他都不转" —— 错误文案也不转)。熔断/指标由调用方按状态码记。 + """ + async with self._lease() as client: + resp = await client.post(self._path(protocol), json=payload, + headers=self._req_headers(session_id, protocol)) + body = resp.content + ctype = resp.headers.get("content-type", "") + if "application/json" in ctype: + try: + return resp.status_code, json.loads(body) + except (ValueError, json.JSONDecodeError): + pass + # 非 JSON(或声明 JSON 但解析失败):把原始文本交回,别在这里编造结构 + return resp.status_code, body.decode("utf-8", "replace") + + @asynccontextmanager + async def raw_stream(self, protocol: str, payload: dict[str, Any], + session_id: str | None = None): + """流式直传:上下文内可直接读 `resp`(状态码 + 字节流)。 + + 为什么用 asynccontextmanager 而不是普通 async generator:调用方需要**先看到 + 状态码**再决定返回 SSE(200)还是上游的 JSON 错误(4xx/5xx)。同时 lease 必须在 + 整个流式期间持有(热重载/关闭不会掐断在途流,与 _stream_once 同一处理)。 + """ + async with self._lease() as client: + async with client.stream("POST", self._path(protocol), json=payload, + headers=self._req_headers(session_id, protocol)) as resp: + yield resp + def _build_payload( self, model: str, @@ -214,7 +357,8 @@ class OpenAICompatAdapter(BaseAdapter): """执行一次非流式 /chat/completions 请求(不含参数修正逻辑)。""" # Bug 4 修复: 用 lease 占用 client,热重载 close 会等本次请求结束才真正关连接 async with self._lease() as client: - resp = await client.post("/chat/completions", json=payload, headers=self._req_headers(session_id)) + resp = await client.post(self._path(PROTOCOL_CHAT), json=payload, + headers=self._req_headers(session_id, PROTOCOL_CHAT)) resp.raise_for_status() data = json_or_raise(resp) @@ -269,7 +413,8 @@ class OpenAICompatAdapter(BaseAdapter): # Bug 4 修复: lease 在整个流式期间持有占用,热重载/关闭不会掐断在途流 async with self._lease() as client: async with client.stream( - "POST", "/chat/completions", json=payload, headers=self._req_headers(session_id) + "POST", self._path(PROTOCOL_CHAT), json=payload, + headers=self._req_headers(session_id, PROTOCOL_CHAT) ) as resp: resp.raise_for_status() async for line in resp.aiter_lines(): @@ -316,7 +461,7 @@ class OpenAICompatAdapter(BaseAdapter): payload["extra_body"] = extra_body async with self._lease() as client: - resp = await client.post(endpoint, json=payload, headers=self._headers) + resp = await client.post(endpoint, json=payload, headers=self._media_headers()) resp.raise_for_status() data = json_or_raise(resp) # 统一:返回 data 数组 @@ -350,7 +495,7 @@ class OpenAICompatAdapter(BaseAdapter): payload[k] = v async with self._lease() as client: - resp = await client.post(endpoint, json=payload, headers=self._headers) + resp = await client.post(endpoint, json=payload, headers=self._media_headers()) resp.raise_for_status() data = json_or_raise(resp) # video_id 用于查询;id/task_id 是任务 id @@ -372,7 +517,7 @@ class OpenAICompatAdapter(BaseAdapter): # 轮询用绝对 URL(poll_base 可能是去掉 /v1 的域名根) url = poll_base.rstrip("/") + poll_path async with self._lease() as client: - resp = await client.get(url, params=params, headers=self._headers) + resp = await client.get(url, params=params, headers=self._media_headers()) resp.raise_for_status() data = json_or_raise(resp) return {"raw": data} diff --git a/app/config.py b/app/config.py index 2c954d6..055e3b0 100644 --- a/app/config.py +++ b/app/config.py @@ -1,7 +1,8 @@ -"""配置加载模块:从 YAML 文件读取供应商、模型、路由、鉴权、运维配置。""" +"""配置加载模块:从 YAML 文件读取供应商、模型、鉴权、运维配置。""" from __future__ import annotations +import logging import os from dataclasses import dataclass, field from pathlib import Path @@ -9,6 +10,8 @@ from typing import Any import yaml +log = logging.getLogger("gateway.config") + class ConfigError(ValueError): """配置结构性错误(缺必填项等)。启动/热重载时抛出,附友好说明。""" @@ -28,6 +31,65 @@ MODEL_PARAM_KEYS: frozenset[str] = frozenset({ }) +# ── 协议线(方言)标准名(第二十七轮「三线直传」) ────────────── +# 老板决策:**入站是什么格式,出站就是什么格式**;网关不做任何跨格式翻译, +# 只在「供应商与模型」里声明每个模型能用哪几条线。 +# +# 这里直接用各家文档 / SDK 里的**端点名**当协议名("标准的来"): +# chat/completions —— OpenAI Chat Completions(POST {base}/chat/completions) +# messages —— Anthropic Messages(POST {base}/messages,认 x-api-key) +# responses —— OpenAI Responses(POST {base}/responses,认 Bearer) +# 好处:配置、日志、前端标签、上游文档四处用的是同一个词,不用查对照表。 +# +# 纯透传的边界: +# - 同一格(入站 = 出站)→ 原样转发,只定点改写 model + 盖模型级强制参数; +# - 不同格(如用 messages 打一个只支持 chat/completions 的模型)→ 入站直接 400 +# 报「该模型不支持 messages 线」,**绝不偷偷翻译**(翻译出来的东西一定有损, +# 而且失败现场离用户十万八千里)。 +PROTOCOL_CHAT = "chat/completions" +PROTOCOL_MESSAGES = "messages" +PROTOCOL_RESPONSES = "responses" +PROTOCOLS: tuple[str, ...] = (PROTOCOL_CHAT, PROTOCOL_MESSAGES, PROTOCOL_RESPONSES) +# 模型没写 protocols 时的默认值:只认 chat/completions。 +# 为什么要默认而不是"全支持":老配置(Agnes / 火山)从未声明过协议, +# 默认全支持等于把「没实测过的线」当成可用,用户打过去只会收到上游 500。 +DEFAULT_PROTOCOLS: tuple[str, ...] = (PROTOCOL_CHAT,) + +# 出站鉴权方式:bearer(Authorization: Bearer xxx)/ x-api-key(Anthropic 系)/ none +AUTH_STYLES: frozenset[str] = frozenset({"bearer", "x-api-key", "none"}) + + +@dataclass +class EndpointSpec: + """某供应商在**一条协议线**上的出站配置(providers..endpoints.)。 + + 为什么放在供应商级而不是模型级:同一条线在同一个供应商下永远是同一个地址与同一套 + 鉴权头(OpenCode Go 三条线共用 https://opencode.ai/zen/go/v1,只是路径和鉴权头不同), + 模型之间只有"能不能走"的差别 —— 那是 models[].protocols 的事。 + """ + + path: str = "" # 相对 base_url 的路径,如 /chat/completions + auth: str = "bearer" # bearer | x-api-key | none + headers: dict[str, str] = field(default_factory=dict) # 该线固定附加头(如 anthropic-version) + + +def _default_endpoints() -> dict[str, EndpointSpec]: + """三条标准线的默认出站配置(未写 providers.*.endpoints 时用)。 + + 默认值就是三家协议的标准形态,写出来只是为了让"默认"可见: + chat/completions → /chat/completions + Bearer + messages → /messages + x-api-key + anthropic-version + responses → /responses + Bearer + """ + return { + PROTOCOL_CHAT: EndpointSpec(path="/chat/completions", auth="bearer"), + PROTOCOL_MESSAGES: EndpointSpec( + path="/messages", auth="x-api-key", + headers={"anthropic-version": "2023-06-01"}), + PROTOCOL_RESPONSES: EndpointSpec(path="/responses", auth="bearer"), + } + + @dataclass class ModelInfo: id: str @@ -36,7 +98,7 @@ class ModelInfo: # 第十三轮: 是否支持「图像输入」(多模态 / 视觉)。 # Agnes 3.0 Flash 文档明确「输入模态:文本、图像 URL」,实测 data:image/png;base64 也能识别; # 其余文本模型(ark-code 等)文档未声明图像输入 → 默认 False。 - # 用途:① /v1/models 暴露给前端标注/告警;② 带图请求走路由别名时把有视觉能力的模型排到前面。 + # 用途:① /v1/models 暴露给前端标注/告警;② 前端据此提示「换个支持视觉的模型」力的模型排到前面。 vision: bool = False # 第二十五轮(模型参数硬约束): 该模型**必须用**的上游参数,发送前强制覆盖客户端传值。 # 背景:OpenCode Go 的 kimi-k2.7-code 只接受 temperature=1,客户端传 0(含一键测速的 @@ -46,6 +108,17 @@ class ModelInfo: # 拉到 0 也不会再 400(值被网关改成上游唯一允许的那个)。 # 键必须在 MODEL_PARAM_KEYS 白名单内,否则 validate_config 直接报错(不静默失效)。 params: dict[str, Any] = field(default_factory=dict) + # 第二十七轮(三线直传): 该模型**能用哪几条协议线**(取值见 PROTOCOLS)。 + # 语义是「入站也用这条线」—— 入站 /v1/messages 打到只声明了 chat/completions 的模型 + # 时,网关直接 400,不替用户翻译(详见 PROTOCOLS 的说明)。 + # 不写 = 默认 ["chat/completions"](老配置行为完全不变)。 + # 实测口径(OpenCode Go 2026-09-18):**文档标注的线一定可用;有些模型实测能走多条线** + # (如 kimi-k3 / qwen 系列 chat + messages 都 200,deepseek 四个模型三线全通), + # 这类按实测把两条线都写上 —— 但只写实测 200 过的,猜测的线不写。 + protocols: list[str] = field(default_factory=lambda: list(DEFAULT_PROTOCOLS)) + + def supports(self, protocol: str) -> bool: + return protocol in self.protocols @dataclass @@ -97,6 +170,17 @@ class ProviderConfig: media: MediaConfig = field(default_factory=MediaConfig) rate_limit_rps: float | None = None # 该供应商全局 QPS 上限(可选) requires_session_header: bool = False # OpenCode Go 等要求每请求动态会话头(x-opencode-session) + # 第二十七轮(三线直传): 该供应商**每条协议线**的出站地址与鉴权头。 + # 未配置时用 _default_endpoints()(三条标准线 + 标准鉴权)—— 所以 Agnes / 火山这类 + # 只走 chat/completions 的供应商一个字段都不用加,行为与以前完全一致。 + endpoints: dict[str, EndpointSpec] = field(default_factory=_default_endpoints) + + def endpoint(self, protocol: str) -> EndpointSpec | None: + """取某条线的出站配置(没有该线 → None,调用方按"供应商不支持这条线"处理)。""" + return self.endpoints.get(protocol) + + def supports_protocol(self, protocol: str) -> bool: + return protocol in self.endpoints # ── 网关入口鉴权 ── @@ -107,24 +191,6 @@ class AuthGate: session_ttl_minutes: int = 1440 # 前端 /auth/login 签发的 session 有效时长(分钟,默认 24h) -# ── 路由别名 + 优先级 fallback ── -@dataclass -class RouteInfo: - """路由别名 → 按优先级排列的 fallback 模型链。 - - models: 第 0 个为最高优先级,调用失败/超时依次切换。 - timeout: 单个模型尝试的超时(秒),None 用 server.timeout。 - retry: 切到下一个模型前,同一模型的重试次数(含首次)。 - rate_limit_rps: 该路由别名整体 QPS 上限(None=不限)。 - """ - - name: str - models: list[str] = field(default_factory=list) - timeout: int | None = None - retry: int = 1 - rate_limit_rps: float | None = None - - @dataclass class CircuitConfig: """供应商熔断器配置。""" @@ -199,7 +265,6 @@ class GatewayConfig: providers: dict[str, ProviderConfig] = field(default_factory=dict) server: ServerConfig = field(default_factory=ServerConfig) auth: AuthGate = field(default_factory=AuthGate) - routes: dict[str, RouteInfo] = field(default_factory=dict) circuit: CircuitConfig = field(default_factory=CircuitConfig) observability: ObservabilityConfig = field(default_factory=ObservabilityConfig) pricing: Pricing = field(default_factory=Pricing) @@ -222,18 +287,29 @@ class GatewayConfig: def find_provider_for_model(self, model_id: str) -> tuple[str, ProviderConfig, ModelInfo] | None: return self._model_index.get(model_id) - def find_route(self, model_id: str) -> RouteInfo | None: - return self.routes.get(model_id) + def find_model_for_protocol( + self, model_id: str, protocol: str + ) -> tuple[str, ProviderConfig, ModelInfo] | None: + """按「模型 + 入站协议线」查供应商;模型不存在或该线不支持 → None。 + + 三线直传的核心闸门:入站是哪条线,就必须在这条线上找到模型,否则 400 —— + **不做任何跨线翻译**(老板决策)。返回 None 的两种情况(未知模型 / 不支持该线) + 由调用方分别给不同文案,别混成一句"模型不存在"。 + """ + found = self._model_index.get(model_id) + if found is None: + return None + pkey, pconfig, minfo = found + if not minfo.supports(protocol): + return None + if not pconfig.supports_protocol(protocol): + return None + return found + + def models_with_protocol(self, protocol: str) -> list[ModelInfo]: + """能走某条线的全部模型(/v1/models 与前端标签用)。""" + return [m for p in self.providers.values() for m in p.models if m.supports(protocol)] - def validate_routes(self) -> list[str]: - """启动校验:路由里的 model ID 是否都存在。返回告警列表(空=全合法)。""" - known = {m["id"] for m in self.all_models()} - warnings: list[str] = [] - for rname, route in self.routes.items(): - for mid in route.models: - if mid not in known: - warnings.append(f"路由 {rname} 引用了未知模型 '{mid}'") - return warnings def all_models(self) -> list[dict[str, Any]]: result = [] @@ -250,20 +326,13 @@ class GatewayConfig: # 前端据此标注「该模型温度已由网关固定」,API 调用方也能提前知道 # 自己传的采样参数会被上游要求覆盖,省掉一轮「传了没生效」的困惑。 "params": dict(m.params or {}), + # 第二十七轮: 暴露该模型能走哪几条协议线(取值见 PROTOCOLS)。 + # 前端「供应商与模型」页据此打格式标签;API 调用方也能提前知道自己 + # 该把请求发到 /v1/chat/completions、/v1/messages 还是 /v1/responses。 + "protocols": list(m.protocols), }) return result - def all_routes(self) -> list[dict[str, Any]]: - out = [] - for rname, r in self.routes.items(): - out.append({ - "name": rname, - "models": r.models, - "timeout": r.timeout, - "retry": r.retry, - }) - return out - def _opt_float(v: Any, field_name: str = "rate_limit_rps") -> float | None: """可选数值解析:None/空串 → None;数字或数字字符串 → float;其它 → ConfigError。 @@ -355,6 +424,82 @@ def _parse_model_params(v: Any, where: str) -> dict[str, Any]: return out +def _parse_protocols(v: Any, where: str) -> list[str]: + """解析模型级 `protocols`(第二十七轮三线直传)。 + + 容错口径(宁可报错,不可静默失效 —— 写错协议名是最容易发生的配置失误, + 而它的表现是"调用该模型永远 400",静默接受等于让用户自己 debug 配置): + - 缺失 / None → 默认 ["chat/completions"](老配置行为不变); + - 字符串 → 当作单元素列表(写 `protocols: messages` 也认); + - 列表 → 逐项校验,必须是 PROTOCOLS 里的标准名; + - 去重保序;空列表 → ConfigError("一个模型一条线都不支持"多半是手滑)。 + """ + if v is None: + return list(DEFAULT_PROTOCOLS) + if isinstance(v, str): + raw_items: list[Any] = [v] + elif isinstance(v, (list, tuple)): + raw_items = list(v) + else: + raise ConfigError( + f"{where} 必须是字符串或列表(可选值:{', '.join(PROTOCOLS)}),得到 {type(v).__name__}: {v!r}") + out: list[str] = [] + for item in raw_items: + name = str(item or "").strip() + if not name: + raise ConfigError(f"{where} 出现空的协议名") + if name not in PROTOCOLS: + raise ConfigError( + f"{where} 里的 '{name}' 不是标准协议名(允许:{', '.join(PROTOCOLS)})") + if name not in out: + out.append(name) + if not out: + raise ConfigError(f"{where} 为空列表 —— 每个模型至少要声明一条协议线") + return out + + +def _parse_endpoints(pval: dict[str, Any], pkey: str) -> dict[str, EndpointSpec]: + """解析供应商级 `endpoints`(每条协议线的出站路径 / 鉴权 / 固定头)。 + + 未配置 → 三条标准线的默认值(见 _default_endpoints)。配置了就**只认配的那些线**: + "没写这条线" 有语义 —— 该供应商没有这条线,模型声明了也没用(调用会明确报错, + 而不是发到一个猜出来的地址上收 404)。 + + 形如: + endpoints: + messages: + path: "/messages" + auth: "x-api-key" + headers: {anthropic-version: "2023-06-01"} + """ + raw = pval.get("endpoints") + if raw is None: + return _default_endpoints() + if not isinstance(raw, dict): + raise ConfigError(f"providers.{pkey}.endpoints 必须是映射,得到 {type(raw).__name__}") + out: dict[str, EndpointSpec] = {} + for pname, spec in raw.items(): + name = str(pname or "").strip() + where = f"providers.{pkey}.endpoints.{name}" + if name not in PROTOCOLS: + raise ConfigError(f"{where} 不是标准协议名(允许:{', '.join(PROTOCOLS)})") + if not isinstance(spec, dict): + raise ConfigError(f"{where} 必须是映射(path/auth/headers),得到 {type(spec).__name__}") + path = str(spec.get("path") or "").strip() + if not path: + raise ConfigError(f"{where}.path 不能为空(相对 base_url 的路径,如 /messages)") + auth = str(spec.get("auth") or "bearer").strip() + if auth not in AUTH_STYLES: + raise ConfigError( + f"{where}.auth 只能是 {'/'.join(sorted(AUTH_STYLES))},得到 {auth!r}") + headers = spec.get("headers") or {} + if not isinstance(headers, dict): + raise ConfigError(f"{where}.headers 必须是映射(如 anthropic-version)") + out[name] = EndpointSpec(path=path, auth=auth, + headers={str(k): str(v) for k, v in headers.items()}) + return out + + def _parse_media(pval: dict[str, Any]) -> MediaConfig: m = pval.get("media", {}) or {} return MediaConfig( @@ -394,7 +539,7 @@ def _parse_download(pval: dict[str, Any]) -> DownloadConfig: def validate_config(cfg: GatewayConfig) -> list[str]: """结构性校验(Bug 12 修复):返回问题列表,空 = 通过。 - 校验供应商必填项与路由链,给出可读的中文说明(而非运行到一半才崩)。 + 校验供应商必填项与模型字段,给出可读的中文说明(而非运行到一半才崩)。 """ issues: list[str] = [] if not cfg.providers: @@ -407,9 +552,13 @@ def validate_config(cfg: GatewayConfig) -> list[str]: for idx, m in enumerate(p.models): if not m.id: issues.append(f"供应商 '{pkey}' 的第 {idx + 1} 个模型缺少 id") - for rname, r in cfg.routes.items(): - if not r.models: - issues.append(f"路由 '{rname}' 的模型链为空") + # 第二十七轮: 模型声明了某条线,但供应商根本没配这条线的出站地址 —— + # 这种配置下调用必然失败(而且会失败在"拼 URL"那一步),必须启动就拦住。 + for proto in m.protocols: + if not p.supports_protocol(proto): + issues.append( + f"供应商 '{pkey}' 的模型 '{m.id}' 声明了协议线 '{proto}'," + f"但该供应商的 endpoints 里没有这条线") # 账号体系已由页面管理(data/users.json,见 app/users.py),配置文件不再承载鉴权。 # 首次启动会创建 admin/admin,不存在「自锁」场景。 return issues @@ -437,24 +586,18 @@ def validate_warnings(cfg: GatewayConfig) -> list[str]: # 供应商缺 Key / 缺模型:上线前就点出来,而不是调用时才 503 / fallback 悄悄失效 for pkey, p in cfg.providers.items(): if not p.models: - warns.append(f"供应商 '{pkey}' 没有任何模型 —— 路由链里指向它的 fallback 项永远不可用") + warns.append(f"供应商 '{pkey}' 没有任何模型 —— 它不会出现在任何一条协议线上,永远不可用") if not p.auth.token and not (p.auth.token_env and os.environ.get(p.auth.token_env)): warns.append(f"供应商 '{pkey}' 未配置 API Key(token 与 token_env 均空/环境变量未设置)" f" —— 调用该供应商的任何模型都会 503") # (账号/密码/API Key 已改由页面管理,弱凭证检查不再适用于配置文件。) - # 路由引用了不存在的模型 id → 该 fallback 项永远不可用(多半是打错字) - known = {m["id"] for m in cfg.all_models()} - for r in cfg.all_routes(): - miss = [mid for mid in r["models"] if mid not in known] - if miss: - warns.append(f"路由 '{r['name']}' 引用了未定义的模型 {miss} —— 这些 fallback 项永远不可用") return warns def load_config(path: str | Path) -> GatewayConfig: """从 YAML 文件加载完整网关配置。 - 结构性错误(缺供应商/缺 base_url/空模型链等)抛 ConfigError,带友好说明。 + 结构性错误(缺供应商/缺 base_url/模型缺 id 等)抛 ConfigError,带友好说明。 """ path = Path(path) if not path.exists(): @@ -484,7 +627,9 @@ def parse_config(raw: dict[str, Any]) -> GatewayConfig: ModelInfo(id=m.get("id", ""), display_name=m.get("display_name", ""), type=m.get("type", "text"), vision=_opt_bool(m.get("vision"), "providers.*.models.*.vision", False), params=_parse_model_params( - m.get("params"), f"providers.{pkey}.models[{idx}].params")) + m.get("params"), f"providers.{pkey}.models[{idx}].params"), + protocols=_parse_protocols( + m.get("protocols"), f"providers.{pkey}.models[{idx}].protocols")) for idx, m in enumerate(pval.get("models", [])) ] providers[pkey] = ProviderConfig( @@ -498,6 +643,7 @@ def parse_config(raw: dict[str, Any]) -> GatewayConfig: rate_limit_rps=_opt_float(pval.get("rate_limit_rps")), requires_session_header=_opt_bool( pval.get("requires_session_header"), "providers.*.requires_session_header", False), + endpoints=_parse_endpoints(pval, pkey), ) srv_raw = raw.get("server", {}) or {} @@ -535,30 +681,10 @@ def parse_config(raw: dict[str, Any]) -> GatewayConfig: auth = AuthGate(session_ttl_minutes=_opt_int(auth_raw.get("session_ttl_minutes"), "auth.session_ttl_minutes", 1440, minimum=1)) - # 路由别名 → 优先级 fallback 链(支持 list 或 dict 两种写法) - routes_raw = raw.get("routes", {}) or {} - routes: dict[str, RouteInfo] = {} - for rname, rval in routes_raw.items(): - if isinstance(rval, dict): - chain: list[str] = [] - for item in (rval.get("models", []) or []): - if isinstance(item, str): - chain.append(item) - elif isinstance(item, dict) and item.get("model"): - chain.append(item["model"]) - routes[rname] = RouteInfo( - name=rname, - models=chain, - # 注意 minimum=0:缺省值是 0(=未设置,下面 `or None` 转 None)。 - # 若写 minimum=1,会让"路由没写 timeout"的配置直接报 ConfigError —— 这正是 - # 第七轮自查抓到的一个自伤 bug(之前只有在配置恰好写了 timeout 时才不炸)。 - timeout=(_opt_int(rval.get("timeout"), f"routes.{rname}.timeout", 0, minimum=0) or None), - retry=_opt_int(rval.get("retry"), f"routes.{rname}.retry", 1, minimum=1), - rate_limit_rps=_opt_float(rval.get("rate_limit_rps")), - ) - else: - chain = [x for x in (rval or []) if isinstance(x, str)] - routes[rname] = RouteInfo(name=rname, models=chain) + # 第二十七轮: 「路由别名 / fallback 链」功能已整体移除。 + # 老配置里残留的 routes 段不再解析,只在日志里提醒一声(不报错,免得旧配置起不来)。 + if raw.get("routes"): + log.warning("配置里的 'routes' 段已废弃(路由别名功能已移除),本次加载忽略该段") # 熔断器 c_raw = raw.get("circuit", {}) or {} @@ -588,7 +714,7 @@ def parse_config(raw: dict[str, Any]) -> GatewayConfig: pricing = Pricing(per_model=per_model) cfg = GatewayConfig( - providers=providers, server=server, auth=auth, routes=routes, + providers=providers, server=server, auth=auth, circuit=circuit, observability=observability, pricing=pricing, download=_parse_download(raw), ) diff --git a/app/main.py b/app/main.py index a9b1625..fcc8124 100644 --- a/app/main.py +++ b/app/main.py @@ -1,9 +1,10 @@ """API 网关主入口:FastAPI 应用(v2.4 全功能版)。 特性: -- 多供应商 OpenAI 兼容 /chat/completions(路由别名按「可用+最快」排序 fallback + 同模型重试 + 熔断) +- 多供应商 OpenAI 兼容 /chat/completions(同模型重试 + 熔断;第二十七轮起三线直传: + /v1/chat/completions、/v1/messages、/v1/responses 各按自家协议原样转发,网关不翻译字段) - 图像 /v1/images/generations、视频 /v1/videos(异步 + 等待模式 + Webhook) -- 客户端 Bearer 鉴权(多用户)+ 全局/路由/供应商级限流 +- 客户端 Bearer 鉴权(多用户)+ 全局/供应商级限流 - 用量指标 /metrics + Prometheus /metrics/prometheus + 访问日志 /logs - 配置热加载、X-Request-ID 贯穿、CORS、gzip、请求体限制 - OpenAI 兼容错误格式 + /v1/models 标准端点 @@ -40,12 +41,14 @@ from starlette.middleware.base import BaseHTTPMiddleware from app.adapters import create_adapter from app.circuit import CircuitBreakerRegistry -from app.config import ConfigError, GatewayConfig, RouteInfo, load_config, validate_warnings +from app.config import (ConfigError, GatewayConfig, load_config, validate_warnings, + PROTOCOL_CHAT, PROTOCOL_MESSAGES, PROTOCOL_RESPONSES, PROTOCOLS) from app.metrics import (Metrics, AccessLog, RateLimiter, BODY_BY_RID, _BODY_MAX, video_poll_push, video_polls, ACCOUNT, ROLE) from app.models import (ChatRequest, ImageRequest, VideoCreateRequestExtended, BatchRequest, - strip_protected, PROTECTED_VIDEO_KEYS, PROTECTED_IMAGE_KEYS) + strip_protected, PROTECTED_VIDEO_KEYS, PROTECTED_IMAGE_KEYS, + normalize_conversation_id) from app.routing import resolve_with_auth, ModelNotFoundError, AuthConfigError from app.sessions import SessionStore from app.users import UsersStore, UserError, key_limit_for_role @@ -142,8 +145,6 @@ def _setup_logging() -> None: root.addHandler(h) -# P0-4: 启动时校验路由模型 ID -_route_warnings = config.validate_routes() # 基础日志先起来,进 lifespan 后按 observability 重配 logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(name)s] %(levelname)s: %(message)s") log = logging.getLogger("api-gateway") @@ -190,10 +191,8 @@ def _warn_security_posture(cfg) -> None: async def lifespan(app: FastAPI): global config _setup_logging() - log.info("API Gateway 启动,已加载 %d 供应商 / %d 路由", len(config.providers), len(config.routes)) + log.info("API Gateway 启动,已加载 %d 供应商", len(config.providers)) _warn_security_posture(config) - for w in _route_warnings: - log.warning("路由校验告警: %s", w) for w in validate_warnings(config): log.warning("配置告警: %s", w) # ⚠ 修复(第五轮,P0 级):这里**绝不能**再注册 SIGTERM/SIGINT 处理器。 @@ -263,14 +262,11 @@ async def _apply_config(new: GatewayConfig, *, source: str = "热重载") -> Non except Exception as e: # 关不掉也要继续换新配置(否则一次异常就卡住整轮) log.warning("关闭旧 adapter 失败(忽略): %s", e) _setup_logging() # Bug 1: 日志级别 / JSON 真正变化 - # Bug O 修复: 换配置后重新校验路由并打印新告警(新增坏路由不再被静默忽略) - for w in new.validate_routes(): - log.warning("路由配置告警: %s", w) for w in validate_warnings(new): log.warning("配置告警: %s", w) _warn_security_posture(new) - log.warning("配置已生效(%s:%d 供应商 / %d 路由 / 熔断=%s)", - source, len(new.providers), len(new.routes), new.circuit.enabled) + log.warning("配置已生效(%s:%d 供应商 / 熔断=%s)", + source, len(new.providers), new.circuit.enabled) async def _hot_reload_loop() -> None: @@ -279,7 +275,7 @@ async def _hot_reload_loop() -> None: Bug 4/E/O 修复: - hot_reload_interval 每轮重新读取(改动即时生效,不再只读一次) - 保留 metrics / 访问日志 / 限流桶 / 熔断状态(不再整体清空) - - 重载后重新计算并打印路由告警 + - 重载后重新计算并打印配置告警 第二十三轮:实际切换逻辑抽到 `_apply_config`(与页面保存共用); server.hot_reload 关掉后循环即停手(不再"关了还在后台重载"); 注意从"关"改回"开"需要重启进程(任务只在 lifespan 里创建)。 @@ -683,36 +679,28 @@ def _get_adapter(model_id: str, timeout: int | None = None, cfg: GatewayConfig | return _state["adapters"][cache_key], pkey, pconfig -# ── 鉴权 + 限流(全局 / 路由 / 供应商级) ── +# ── 鉴权 + 限流(全局 / 供应商级;路由级限流随路由功能一并移除) ── _bearer = HTTPBearer(auto_error=False) _sess = SessionStore() # 前端登录会话(/auth/login 换取 sess- token,见 app/sessions.py) -def _effective_rps(cfg, route: RouteInfo | None, pkey: str | None) -> float | None: - """限流优先级:路由级 > 供应商级 > 全局。""" - if route and route.rate_limit_rps: - return route.rate_limit_rps +def _effective_rps(cfg, pkey: str | None) -> float | None: + """限流优先级:供应商级 > 全局。""" if pkey and cfg.providers.get(pkey) and cfg.providers[pkey].rate_limit_rps: return cfg.providers[pkey].rate_limit_rps return cfg.server.rate_limit_rps -def _enforce_rate_limit(cfg, request, route: RouteInfo | None, pkey: str | None, model_id: str) -> None: - """按 路由级 > 供应商级 > 全局 三级限流(Bug 1 修复)。 +def _enforce_rate_limit(cfg, request, pkey: str | None, model_id: str) -> None: + """按 供应商级 > 全局 两级限流(Bug 1 修复)。 - 不同维度用不同 limiter 桶(scope 前缀),避免互相污染: - - 路由级:bucket=route:,限流则 429(整条路由打满) - - 供应商级:bucket=prov:,限流则 429 - 全局级在 require_auth 里已按 key 处理。 + 供应商级用 bucket=prov:(与媒体端点共桶),限流则 429; + 全局级在 require_auth 里已按 key 处理。路由级限流随路由功能一并移除。 """ - rps = _effective_rps(cfg, route, pkey) + rps = _effective_rps(cfg, pkey) if not rps: return - # 路由级优先 - if route and route.rate_limit_rps: - if not _limiter.allow(f"route:{route.name}", route.rate_limit_rps): - raise HTTPException(429, f"路由 {route.name} 触发限流({route.rate_limit_rps} rps)") - elif pkey and (cfg.providers.get(pkey) or None) and cfg.providers[pkey].rate_limit_rps: + if pkey and (cfg.providers.get(pkey) or None) and cfg.providers[pkey].rate_limit_rps: if not _limiter.allow(f"prov:{pkey}", cfg.providers[pkey].rate_limit_rps): raise HTTPException(429, f"供应商 {pkey} 触发限流({cfg.providers[pkey].rate_limit_rps} rps)") @@ -944,6 +932,12 @@ async def require_auth( cfg = _cfg() _set_identity(request, "anonymous", "member") key = credentials.credentials if credentials else "" + # 第二十七轮(三线直传): Anthropic 系客户端(Claude Code / anthropic-sdk)默认把 + # 凭证放在 `x-api-key` 而不是 `Authorization: Bearer` —— 既然网关对外提供 + # /v1/messages,就必须认这个头,否则"格式对了但连鉴权都过不去"。 + # 优先级:Authorization 优先(显式标准头),没有才回落到 x-api-key。 + if not key: + key = (request.headers.get("x-api-key") or "").strip() identity = _identity_from_credentials(key) if identity is None: # 第二十四轮: Key 本身是有效的、只是所属账号被管理员停用 → 回 403 明确告知原因, @@ -989,60 +983,25 @@ async def require_auth_or_open( raise HTTPException(401, "无效或缺失的账号 / API Key", headers={"WWW-Authenticate": "Bearer"}) -def _model_available_and_speed(cfg: GatewayConfig, model_id: str) -> tuple[bool, float | None]: - """该文本模型「当前是否可用」+「最近平均延迟」。 +def _require_protocol_model(cfg: GatewayConfig, model: str, protocol: str) -> str: + """校验「模型存在 + 声明了这条协议线」,返回模型 id(第二十七轮三线直传的闸门)。 - 可用 = 所属供应商配了凭证 **且** 熔断未处于 OPEN(HALF_OPEN 视为可用,它本就是恢复试探)。 - 延迟取 metrics 里该模型最近若干次的均值;没调用过则 None(上层按「未测速」处理)。 + 路由别名 / fallback 链已在第二十七轮整体移除:入站 model 必须是真实模型 id, + 网关不再"帮你挑一个"、也不会跨线降级。为什么必须在这里挡:实测(OpenCode Go + 2026-09-18)把 chat 线的模型打到 /messages 上,上游回的是 **500 Internal server + error**;把 messages 线的模型打到 /responses 上,回的是 401 "not supported for + format openai"。两种都不像"你选错了格式",与其让用户对着这种错误猜,不如网关 + 直接 400 说清楚——候选支持哪几条线也一并报出来。 """ - found = cfg.find_provider_for_model(model_id) - if not found: - return (False, None) - pkey, pconfig, _minfo = found - try: - token_ok = bool(pconfig.auth.resolve_token()) - except Exception: - token_ok = False - state = _circuits.get(pkey).status().get("state", "closed") - return (token_ok and state != "open", _metrics.recent_avg_latency(model_id)) - - -def _order_chain_by_speed(cfg: GatewayConfig, chain: list[str]) -> list[str]: - """第十一轮 #4:wjd 这类路由**不再按配置优先级**,改为 - - ① 先剔除「当前不可用」的模型(供应商没凭证 / 熔断 OPEN)—— 排到最后; - ② 可用模型按最近平均延迟升序,最快的先调用; - ③ 没测到延迟的可用模型给「探测位」:排在有数据的之后、不可用的之前, - 这样新模型仍会被轮到去学习它的速度,但不会盖过已知的快模型; - ④ 冷启动(整条链都没样本)时**原样返回**,等价于退回配置的优先级顺序。 - """ - decorated = [] - for idx, mid in enumerate(chain): - avail, lat = _model_available_and_speed(cfg, mid) - decorated.append((mid, avail, lat, idx)) - if not any(lat is not None for _, _a, lat, _i in decorated): - return chain # 还没任何速度数据 → 保持配置顺序,别乱猜 - known = [lat for _m, a, lat, _i in decorated if a and lat is not None] - probe_rank = (max(known) + 1.0) if known else 0.0 - - def _key(item): - _mid, avail, lat, idx = item - # 可用优先;可用里按速度(未测速的用 probe_rank 兜到已测慢模型之后);同权重保持原序 - speed = (probe_rank if lat is None else lat) if avail else 0.0 - return (0 if avail else 1, speed, idx) - - decorated.sort(key=_key) - return [mid for mid, _a, _lat, _i in decorated] - - -def _route_chain(cfg: GatewayConfig, requested: str) -> tuple[list[str], RouteInfo | None]: - route = cfg.find_route(requested) - chain = list(route.models) if (route and route.models) else [requested] - # 第十一轮 #4:路由别名(如 wjd)按「可用 + 速度」动态排序,不再走静态优先级; - # 直接填真实模型 id(route is None)时不动,保持单次调用的确定性。 - if route is not None: - chain = _order_chain_by_speed(cfg, chain) - return chain, route + found = cfg.find_provider_for_model(model) + if found is None: + raise HTTPException(400, f"未知模型: {model}") + if cfg.find_model_for_protocol(model, protocol) is None: + raise HTTPException( + 400, + f"模型 {model} 不支持 {protocol} 线(它支持:{'、'.join(found[2].protocols) or '无'})。" + f"网关不做跨格式转换,请改用模型支持的线,或换一个支持 {protocol} 的模型。") + return model # ── 第十三轮: 视觉(带图)请求的模型选择 ───────────────────────── @@ -1068,53 +1027,20 @@ def messages_have_images(messages: list[dict]) -> bool: return False -def _chain_has_vision(cfg: GatewayConfig, chain: list[str]) -> bool: - for mid in chain: - found = cfg.find_provider_for_model(mid) - if found and found[2].vision: - return True - return False - - -def _vision_first(cfg: GatewayConfig, chain: list[str]) -> list[str]: - """带图请求:把有视觉能力的模型提到链首(保持各自原有相对顺序)。 - - 背景:路由别名(如 wjd)的链首常是纯文本模型(火山 ark-code)。带图请求打到 - 纯文本模型上,上游要么忽略图片(用户以为"看图了"其实没看),要么直接报参数错误 —— - 两种情况都很难自查。这里只在**链上确实有视觉模型**时重排;一个都没有就原样返回, - 让上游如实报错(网关不替用户隐藏能力缺失)。 - """ - def _vision(mid: str) -> bool: - found = cfg.find_provider_for_model(mid) - return bool(found and found[2].vision) - - visual = [m for m in chain if _vision(m)] - if not visual: - return chain - return visual + [m for m in chain if not _vision(m)] - - def _resolve_media_model(cfg, model: str, media_type: str): """Bug 4: 解析媒体模型。 - model 是真实模型 ID 且类型匹配(或 text)→ 直接 resolve - - model 是路由别名 → 遍历路由链找第一个 type==media_type 的模型并 resolve - 找不到 → 返回 None + + 第二十七轮: 路由别名(如 wjd)解析分支已随路由功能整体移除 —— 媒体端点也只认真实模型 id。 """ found = cfg.find_provider_for_model(model) if found: pkey, pconfig, minfo = found if minfo.type == media_type or media_type == "text": return pkey, pconfig, minfo, pconfig.auth.resolve_token(), model - # 当作路由别名解析 - route = cfg.find_route(model) - if route: - for mid in route.models: - f2 = cfg.find_provider_for_model(mid) - if f2 and f2[2].type == media_type: - pkey, pconfig, minfo = f2 - return pkey, pconfig, minfo, pconfig.auth.resolve_token(), mid - return None + def _resp_text(res: dict) -> str: @@ -1134,6 +1060,27 @@ def _resp_text(res: dict) -> str: reasoning = str(msg.get("reasoning_content") or "") if reasoning.strip(): return "[推理] " + reasoning + # 第二十七轮: messages / responses 线的响应结构完全不同(没有 choices)。 + # 访问日志要能看到"这次到底回了什么",所以顺手认一下这两家的正文位置 + # —— 只影响日志展示,**不改任何响应内容**。 + # messages: content=[{type:text,text:...}] + # responses: output_text(部分实现直接给)/ output[].content[].text + blocks = res.get("content") + if isinstance(blocks, list): + text = "".join(str(b.get("text") or "") for b in blocks if isinstance(b, dict)) + if text.strip(): + return text + if isinstance(res.get("output_text"), str) and res["output_text"].strip(): + return res["output_text"] + out = res.get("output") + if isinstance(out, list): + parts = [] + for item in out: + for b in (item.get("content") or []) if isinstance(item, dict) else []: + if isinstance(b, dict) and b.get("text"): + parts.append(str(b["text"])) + if parts: + return "".join(parts) except Exception: pass return "" @@ -1167,16 +1114,15 @@ def _attach_cost(resp: dict, model: str, usage: dict, cfg) -> None: def _models_payload(): cfg = _cfg() # 优化: 目录按配置 mtime 缓存(热重载改配置才重建;静默期每次请求直接命中,省掉 - # all_models()/all_routes() 的重复遍历与对象构建)。cfg 引用同一 config 对象, + # all_models() 的重复遍历与对象构建)。cfg 引用同一 config 对象, # 但在 hot_reload 时 mtime 会变 → 缓存自然失效。 mtime = _state.get("mtime") _cached = _state.get("_models_cache") if _cached is not None and _cached[0] == mtime: return _cached[1] - # 第六轮: 统一 data 里的 object 取值 —— 此前路由别名是 object="route",严格按 - # `object == "model"` 过滤的 OpenAI 客户端会解析出错。路由别名本身是**可调用的 model id**, - # 所以保留在 data 里但改用 object="model"(用 type="route" 区分), - # 同时补齐 OpenAI 标准的 created / owned_by 字段。 + # 第六轮: 统一 data 里的 object 取值(一律 "model"),并补齐 OpenAI 标准的 + # created / owned_by 字段 —— 严格按 `object == "model"` 过滤的客户端才认。 + # 第二十七轮: 路由别名(type="route")条目随路由功能整体移除,这里只剩真实模型。 created = int(_state.get("mtime") or 0) data = [] for m in cfg.all_models(): @@ -1188,16 +1134,15 @@ def _models_payload(): "vision": bool(m.get("vision")), # 第二十五轮: 暴露模型级强制参数(如 {"temperature": 1})。前端据此标挂锁, # API 调用方也能提前知道自己传的采样参数会被上游要求覆盖。 - "params": dict(m.get("params") or {})}) - for r in cfg.all_routes(): - data.append({"id": r["name"], "object": "model", "created": created, - "owned_by": "gateway", "type": "route", "models": r["models"], - # 路由只要链上有任一视觉模型就算 vision:带图请求会被排到那个模型上(见 _vision_first) - "vision": _chain_has_vision(cfg, r["models"])}) + "params": dict(m.get("params") or {}), + # 第二十七轮(三线直传): 该模型能走哪几条协议线 + # (chat/completions | messages | responses)。前端据此打格式标签, + # 调用方据此决定把请求发到 /v1/chat/completions、/v1/messages 还是 + # /v1/responses —— 网关不做任何跨线翻译,写错线直接 400。 + "protocols": list(m.get("protocols") or [PROTOCOL_CHAT])}) payload = { "object": "list", "data": data, - "routes": cfg.all_routes(), # 本轮修复: 不再回显 base_url(内部拓扑信息,未开鉴权时会被任意访问者读到) "providers": {k: {"name": v.name, "model_count": len(v.models)} for k, v in cfg.providers.items()}, @@ -1216,10 +1161,335 @@ async def v1_models(request: Request, _auth=Depends(require_auth)): return _models_payload() +# ── 第二十七轮「三线直传」:/v1/messages 与 /v1/responses ────────────── +# 老板决策:**入站是什么格式,出站就是什么格式**;网关不翻译任何协议内容,只在 +# 「供应商与模型」里声明每个模型能用哪几条协议线(models[].protocols)。 +# +# 与 /v1/chat/completions 的关系: +# chat/completions 线走**专用路径**(ChatRequest 校验 + 适配器按字段组装 payload), +# 这两条新线走**通用直传**(原始 JSON 原样进去、上游响应原样回来)—— 原因是 +# 这两家协议的结构(内容块、tool 表达、SSE 事件类型)差别很大,重建一遍字段 +# 就是"翻译",会引入损耗,正好违背老板要的"不转"。 +# +# 网关在直传里仍然做的(不算转格式,是网关的立身之本): +# ① 挑供应商与协议线:模型不支持这条线 → 400,绝不偷偷翻译; +# ② 定点改写 model(别名 → 真实 id); +# ③ 盖模型级强制参数(models[].params,如 kimi-k2.7-code 只接受 temperature=1); +# ④ 注入该线的鉴权头 + x-opencode-session(入站 conversation_id 派生成按账号隔离的会话 id); +# ⑤ 熔断 / 限流 / 记账(usage 三种形状各解析一次,只读)/ 访问日志; +# ⑥ 补网关壳字段(served_model/provider/provider_name/protocol/x_request_id)。 +# 明确**不做**:字段改名、内容块互转、工具调用互转、SSE 事件互转、错误文案翻译。 + +def _protocol_mismatch(minfo, model_id: str, protocol: str) -> str: + """模型不支持入站协议线时的 400 文案(说清"它支持什么",别让用户猜)。""" + supported = "、".join(minfo.protocols) or "(未声明)" + others = "、".join(p for p in PROTOCOLS if p != protocol) + return (f"模型 {model_id} 不支持 {protocol} 格式(它支持:{supported})。" + f"网关不做跨格式转换:请改用 {supported} 对应的端点,或换一个支持 {protocol} 的模型" + f"(如 {others} 线的模型)。") + + +def _usage_counts(protocol: str, data: Any) -> dict[str, int]: + """把三家协议的 usage 归一成 (prompt_tokens, completion_tokens) 供记账。 + + 这是**只读解析**:上游响应体原样交给客户端,这里只是为了指标/成本不丢 + —— chats 用 prompt_tokens/completion_tokens,messages 与 responses 都用 + input_tokens/output_tokens(不解析的话,两条新线的调用在统计里会全是 0)。 + """ + if not isinstance(data, dict): + return {"prompt_tokens": 0, "completion_tokens": 0} + u = data.get("usage") + if not isinstance(u, dict): + return {"prompt_tokens": 0, "completion_tokens": 0} + if protocol in (PROTOCOL_MESSAGES, PROTOCOL_RESPONSES): + return {"prompt_tokens": int(u.get("input_tokens") or 0), + "completion_tokens": int(u.get("output_tokens") or 0)} + return {"prompt_tokens": int(u.get("prompt_tokens") or 0), + "completion_tokens": int(u.get("completion_tokens") or 0)} + + +class _UsageScanner: + """流式直传的 usage 扫描器(只读,不改一个字节)。 + + 流式响应的 usage 藏在 SSE 事件里,且三家位置不同: + chat/completions → `data: {"usage": {...}}`(需 include_usage) + messages → event `message_start`(input_tokens)+ `message_delta`(output_tokens) + responses → event `response.completed`(response.usage) + 这里按行喂入,识别 data: 行里的 JSON,并把常见的三种嵌套位置都看一眼, + 取到就更新(不依赖事件顺序 —— 两个方向各自覆盖,缺一个也不影响另一个)。 + """ + + def __init__(self, protocol: str): + self.protocol = protocol + self.prompt = 0 + self.completion = 0 + self.seen = False + self._buf = "" + + def feed(self, chunk: bytes) -> None: + try: + text = chunk.decode("utf-8") + except UnicodeDecodeError: + return # 跨 chunk 的多字节字符会在下一个 chunk 里补齐(这里丢掉半截即可) + self._buf += text + while "\n" in self._buf: + line, self._buf = self._buf.split("\n", 1) + self._line(line.strip()) + + def _line(self, line: str) -> None: + if not line.startswith("data:"): + return + payload = line[5:].strip() + if not payload or payload == "[DONE]": + return + try: + obj = json.loads(payload) + except (ValueError, json.JSONDecodeError): + return + if not isinstance(obj, dict): + return + for cand in (obj, obj.get("message"), obj.get("response")): + if not isinstance(cand, dict): + continue + u = cand.get("usage") + if not isinstance(u, dict): + continue + inp = u.get("input_tokens", u.get("prompt_tokens")) + out = u.get("output_tokens", u.get("completion_tokens")) + if inp is not None: + self.prompt = int(inp) + self.seen = True + if out is not None: + self.completion = int(out) + self.seen = True + + def counts(self) -> dict[str, int]: + return {"prompt_tokens": self.prompt, "completion_tokens": self.completion} + + +def _resolve_passthrough(cfg: GatewayConfig, body: dict, protocol: str): + """直传入口的公共解析:取 model + 校验协议线 + 抽 conversation_id。 + + 返回 (model_id, pkey, pconfig, minfo, conversation_id, clean_body); + clean_body 已摘掉网关内部字段(conversation_id),其余原样。 + """ + model_in = str(body.get("model") or "").strip() + if not model_in: + raise HTTPException(400, "缺少 model 字段") + clean = dict(body) + try: + conv = normalize_conversation_id(clean.pop("conversation_id", None)) + except ValueError as e: + raise HTTPException(400, f"conversation_id {e}") + found = cfg.find_provider_for_model(model_in) + if found is None: + raise HTTPException(400, f"未知模型: {model_in}") + pkey, pconfig, minfo = found + if not minfo.supports(protocol): + raise HTTPException(400, _protocol_mismatch(minfo, model_in, protocol)) + if not pconfig.supports_protocol(protocol): + raise HTTPException(400, f"供应商 {pconfig.name} 未配置 {protocol} 线") + return model_in, pkey, pconfig, minfo, conv, clean + + +async def _passthrough(request: Request, protocol: str): + """三线直传的通用实现(当前服务 /v1/messages 与 /v1/responses)。""" + cfg = _cfg() + rid = getattr(request.state, "rid", "-") + caller = getattr(request.state, "caller", "anonymous") + raw = await request.body() + if not raw: + raise HTTPException(400, "请求体为空") + try: + body = json.loads(raw) + except (ValueError, json.JSONDecodeError): + raise HTTPException(400, "请求体不是合法 JSON") + if not isinstance(body, dict): + raise HTTPException(400, "请求体必须是 JSON 对象") + + model_in, pkey, pconfig, _minfo, conv, clean = _resolve_passthrough(cfg, body, protocol) + stream = bool(clean.get("stream")) + + try: + adapter, pkey, pconfig = _get_adapter(model_in, cfg.server.timeout, cfg) + except (ModelNotFoundError, AuthConfigError) as e: + raise HTTPException(502, str(e)) + cb = _circuits.get(pkey) + if not cb.allow(): + raise HTTPException(503, f"供应商 {pconfig.name} 熔断中,请稍后重试") + try: + _enforce_rate_limit(cfg, request, pkey, model_in) + except HTTPException: + cb.release_probe() # 请求没发出去 → 归还半开试探令牌 + raise + + # 定点改写:model 换成真实 id(别名场景)+ 盖模型级强制参数;其余字段一个不动 + payload = adapter.prepare_payload(model_in, clean) + session = _upstream_session(conv, caller) + + if stream: + return await _passthrough_stream(request, rid, caller, protocol, adapter, cb, + pkey, pconfig, model_in, payload, session, cfg) + + t0 = time.time() + try: + status, data = await adapter.raw_json(protocol, payload, session) + except Exception as e: + cb.record(False) + _metrics.record_call(model_in, pkey, ok=False) + _metrics.record_latency(model_in, time.time() - t0) + log.error("[rid=%s] %s 直传失败 [%s]: %s", rid, protocol, model_in, e) + raise HTTPException(502, f"上游调用失败: {e}") + + ok = 200 <= status < 300 + # 熔断口径(与真实链路一致):4xx 是**客户端 body 的问题**(直传下尤其如此, + # body 就是客户端自己写的),不该把供应商判成坏了;只有 5xx 才算上游故障。 + if ok or status >= 500: + cb.record(ok) + else: + cb.release_probe() + _metrics.record_call(model_in, pkey, ok=ok) + _metrics.record_latency(model_in, time.time() - t0) + usage = _usage_counts(protocol, data) + if ok: + _metrics.record_tokens(usage["prompt_tokens"], usage["completion_tokens"]) + _access.add({"method": "POST", "kind": "chat", "path": f"/v1/{protocol}", "rid": rid, + "model": model_in, "served": model_in, "provider": pkey, + "caller": caller, "ok": ok, "stream": False, + "response": _logsafe(_resp_text(data) if ok else _upstream_detail( + json.dumps(data, ensure_ascii=False) if isinstance(data, dict) else str(data)), 2000)}) + + if not ok: + # 直传口径:上游的错误体**原样**回给客户端(错误文案也不翻译),只加一个 + # 网关请求号,方便拿它去日志里对账。 + content = data if isinstance(data, dict) else {"error": {"message": str(data)[:1000]}} + return JSONResponse(status_code=status, content=content, + headers={"x-request-id": rid}) + if not isinstance(data, dict): + return Response(content=str(data), media_type="application/json", + headers={"x-request-id": rid}) + resp = dict(data) + if usage["prompt_tokens"] or usage["completion_tokens"]: + _attach_cost(resp, model_in, usage, cfg) + resp["served_model"] = model_in + resp["provider"] = pkey + resp["provider_name"] = pconfig.name + resp["protocol"] = protocol + resp["x_request_id"] = rid + return JSONResponse(status_code=status, content=resp, headers={"x-request-id": rid}) + + +async def _passthrough_stream(request, rid, caller, protocol, adapter, cb, + pkey, pconfig, model_in, payload, session, cfg): + """直传的流式分支:**原样转发上游 SSE 字节**(不解析事件、不重组帧)。 + + 为什么要先 `__aenter__` 再建 StreamingResponse:客户端需要拿到**真实状态码**。 + 上游拒绝一个 stream 请求(400/401)时,若把它塞进 200 的 SSE 流里,SDK 会以为 + 请求成功了、只是没内容 —— 这与直传"如实反映上游"的初衷相反。 + """ + cm = adapter.raw_stream(protocol, payload, session) + resp = await cm.__aenter__() + headers = {"x-request-id": rid} + if resp.status_code >= 400: + raw = await resp.aread() + ctype = resp.headers.get("content-type", "application/json") + await cm.__aexit__(None, None, None) + if resp.status_code >= 500: + cb.record(False) + else: + cb.release_probe() + _metrics.record_call(model_in, pkey, ok=False) + _access.add({"method": "POST", "kind": "chat", "path": f"/v1/{protocol}", "rid": rid, + "model": model_in, "served": model_in, "provider": pkey, "caller": caller, + "ok": False, "stream": True, + "response": _logsafe(_upstream_detail(raw.decode("utf-8", "replace")), 2000)}) + return Response(content=raw, status_code=resp.status_code, media_type=ctype, headers=headers) + + scanner = _UsageScanner(protocol) + t0 = time.time() + media_type = resp.headers.get("content-type", "text/event-stream") + + async def event_stream(): + ok = True + try: + async for chunk in resp.aiter_bytes(): + scanner.feed(chunk) # 只读扫描 usage;字节原样 YIELD + yield chunk + except Exception as e: + ok = False + log.error("[rid=%s] %s 直传流中断 [%s]: %s", rid, protocol, model_in, e) + raise + finally: + await cm.__aexit__(None, None, None) + cb.record(ok) + _metrics.record_call(model_in, pkey, ok=ok) + _metrics.record_latency(model_in, time.time() - t0) + counts = scanner.counts() + if ok and scanner.seen: + _metrics.record_tokens(counts["prompt_tokens"], counts["completion_tokens"]) + _access.add({"method": "POST", "kind": "chat", "path": f"/v1/{protocol}", "rid": rid, + "model": model_in, "served": model_in, "provider": pkey, + "caller": caller, "ok": ok, "stream": True, "response": ""}) + + return StreamingResponse(event_stream(), media_type=media_type, headers=headers) + + +@app.post("/v1/messages") +async def anthropic_messages(request: Request, _auth=Depends(require_auth)): + """Anthropic Messages 线(Claude Code / anthropic-sdk 的默认端点)。 + + 入站 = 出站 = `messages`:body 原样转发到该供应商的 messages 线(鉴权头按该线 + 配置,OpenCode Go 实测认 `x-api-key` + `anthropic-version`),响应(含 SSE)原样回。 + 鉴权:`Authorization: Bearer <网关 Key>` 与 `x-api-key: <网关 Key>` 都认(见 require_auth)。 + """ + return await _passthrough(request, PROTOCOL_MESSAGES) + + +@app.post("/v1/responses") +async def openai_responses(request: Request, _auth=Depends(require_auth)): + """OpenAI Responses 线(Codex / openai-sdk 的 responses 端点)。 + + 入站 = 出站 = `responses`:body 原样转发到该供应商的 responses 线(Bearer 鉴权), + 响应(含 SSE)原样回。模型不支持该线时 400(不做跨格式翻译)。 + """ + return await _passthrough(request, PROTOCOL_RESPONSES) + + # ── 一键测速(第二十七轮: 主动探测全部文本模型延迟)─────────────────── # 复用主调用链的 _get_adapter / adapter.chat(同一适配器、同一供应商 Key), # 对每个文本模型发一条 1-token 的 "hi" 探针并计时,返回「延迟升序」结果。 # 注意:只读展示,不把探针写进 _metrics,避免一条小请求反过来改生产路由排序。 +def _probe_protocol(minfo) -> str: + """探针走哪条线:优先 chat/completions,其次 messages,最后 responses。 + + 第二十七轮: 不是每个模型都能走 chat/completions(实测 grok-4.6 / gpt-5.6-luna + 只认 responses;minimax-m2.7 只认 messages)。一键测速要对**每个**模型都能出结果, + 所以按模型自己声明的线挑一条 —— 挑不到就说明配置有问题(返回空串,调用方给文案)。 + """ + for p in (PROTOCOL_CHAT, PROTOCOL_MESSAGES, PROTOCOL_RESPONSES): + if minfo.supports(p): + return p + return "" + + +def _probe_payload(protocol: str, model_id: str) -> dict: + """探针请求体:每条线一份**最小合法**结构(不翻译,只是各写各的)。 + + - chat/completions: messages + max_tokens + - messages: max_tokens 必填 + messages + - responses: input + max_output_tokens + max_tokens 固定用 16:部分上游对 max_tokens<=2 回 400("must be greater than 2"), + 而 16 足够便宜(探针成本可忽略),也够推理模型吐出一段内容。 + """ + if protocol == PROTOCOL_MESSAGES: + return {"model": model_id, "max_tokens": 16, + "messages": [{"role": "user", "content": "hi"}]} + if protocol == PROTOCOL_RESPONSES: + return {"model": model_id, "input": "hi", "max_output_tokens": 16} + return {"model": model_id, "messages": [{"role": "user", "content": "hi"}], + "max_tokens": 16, "stream": False} + + @app.post("/v1/models/speedtest") async def models_speedtest(request: Request, _auth=Depends(require_auth)): cfg = _cfg() @@ -1243,7 +1513,18 @@ async def models_speedtest(request: Request, _auth=Depends(require_auth)): if only and m["id"] != only: continue row: dict = {"model": m["id"], "display_name": m.get("display_name"), - "provider": m.get("provider_name") or m.get("provider") or "—"} + "provider": m.get("provider_name") or m.get("provider") or "—", + # 第二十七轮: 探针走哪条线也回给前端(模型列表里能对照格式标签) + "protocol": ""} + # 第二十七轮: 按模型自己声明的协议线探测 —— 三条线共用同一个探针实现, + # 否则 responses-only / messages-only 的模型会被误报成"调用失败"。 + found = cfg.find_provider_for_model(m["id"]) + protocol = _probe_protocol(found[2]) if found else "" + if not protocol: + row.update({"ok": False, "latency_ms": None, "error": "该模型未声明任何协议线"}) + results.append(row) + continue + row["protocol"] = protocol try: adapter, pkey, pconfig = _get_adapter(m["id"], int(probe_timeout), cfg) except ModelNotFoundError as e: @@ -1265,12 +1546,18 @@ async def models_speedtest(request: Request, _auth=Depends(require_auth)): # 对 429/5xx 重试(退避同真实链路),4xx/超时不重试。 for attempt in range(1, 4): # 最多 3 次(含首次) try: - await adapter.chat(model=m["id"], messages=[{"role": "user", "content": "hi"}], - temperature=0, max_tokens=16, stop=None) - # max_tokens 固定 >2:部分上游对 max_tokens<=2 回 400(要求 >2),故探针用 16 而不是 1,保证各供应商上游都认。 - # ("max_tokens must be greater than 2") - ok = True - break + # 直传原语:入站 body 原样上去、上游响应原样回来。探针只关心"通不通", + # 所以不解析响应体,只看状态码(4xx/5xx 都算失败并带上游原文)。 + status, data = await adapter.raw_json(protocol, _probe_payload(protocol, m["id"])) + if 200 <= status < 300: + ok = True + break + err = _upstream_detail(json.dumps(data, ensure_ascii=False) + if isinstance(data, dict) else str(data)) or f"上游返回 HTTP {status}" + if attempt < 3 and (status == 429 or status >= 500): + await asyncio.sleep(0.3 * attempt) + else: + break except Exception as e: err = _speedtest_error_text(e, probe_timeout) if attempt < 3 and _speedtest_retryable(e): @@ -1306,26 +1593,20 @@ async def chat_completions(request: Request, req: ChatRequest, _auth=Depends(req messages = [m.to_upstream() for m in req.messages] tools = [t.model_dump(exclude_none=True) for t in req.tools] if req.tools else None - chain, route = _route_chain(cfg, req.model) - if not chain: - raise HTTPException(400, f"路由 {req.model} 未配置有效模型链") - # Bug 15 修复: 未知模型/路由应作为客户端错误 400 直接拒绝, - # 而不是等到 fallback 逐个尝试后才以 502(上游错误)误导性地返回。 - if route is None and cfg.find_provider_for_model(req.model) is None: - raise HTTPException(400, f"未知模型或路由: {req.model}") + # Bug 15 修复: 未知模型应作为客户端错误 400 直接拒绝, + # 而不是等到调用失败后才以 502(上游错误)误导性地返回。 + # 第二十七轮(三线直传): 这条入口就是 chat/completions 线 —— 模型必须声明该线, + # 否则 400(不把请求发去上游吃 500)。路由别名 / fallback 链已整体移除。 + model_id = _require_protocol_model(cfg, req.model, PROTOCOL_CHAT) - # 第十三轮: 带图的视觉请求把有视觉能力的模型排到链首(仅路由别名会重排; - # 直接指定真实模型 id 时不改用户的选择,只记一条日志)。 - with_images = messages_have_images(messages) - if with_images and route is not None: - reordered = _vision_first(cfg, chain) - if reordered != chain: - log.info("[rid=%s] 带图请求,链序调整为视觉优先: %s", rid, " → ".join(reordered)) - chain = reordered + # 第十三轮: 带图请求打到没标视觉能力的模型上,上游要么忽略图片、要么报参数错 —— + # 记一条日志便于排查。第二十七轮: 不再重排模型(只有一个模型,重排=替用户换模型)。 + if messages_have_images(messages) and not cfg.find_provider_for_model(model_id)[2].vision: + log.info("[rid=%s] 带图请求,但模型 %s 未标记视觉能力", rid, model_id) if req.stream: - return _stream_response(request, req, chain, route, messages, tools) - return await _chat_fallback(cfg, rid, caller, request, req, chain, route, messages, tools) + return _stream_response(request, req, model_id, messages, tools) + return await _chat_once(cfg, rid, caller, request, req, model_id, messages, tools) def _upstream_session(conversation_id: str | None, account: str) -> str | None: @@ -1347,256 +1628,222 @@ def _upstream_session(conversation_id: str | None, account: str) -> str | None: return hashlib.sha256(f"{account}:{conversation_id}".encode("utf-8")).hexdigest()[:32] -async def _chat_fallback(cfg, rid, caller, request, req, chain, route, messages, tools): - """非流式:按优先级逐个尝试 + 熔断 + 重试。""" - errors: list[dict] = [] +async def _chat_once(cfg, rid, caller, request, req, model_id, messages, tools): + """非流式:单个模型的调用 + 熔断 + 重试。 + + 第二十七轮: 路由别名 / fallback 链已整体移除 —— 这里只服务一个模型,失败不再"换一个 + 再试",而是把上游错误如实报给客户端(同模型重试仍受 server.retry 控制)。 + """ tool_choice = req.tool_choice + timeout = cfg.server.timeout + try: + adapter, pkey, pconfig = _get_adapter(model_id, timeout, cfg) + except ModelNotFoundError as e: + raise HTTPException(400, f"未知模型: {model_id}") from e + except AuthConfigError as e: + # 供应商没配 Key 是网关侧配置问题:503(别报 502 让用户去查上游) + raise HTTPException(503, f"供应商 {e.provider_key} 未配置可用的 API Key({e.detail})") from e - for model_id in chain: - # Bug 2 修复: route 存在时整条链(含链首)都用 route.retry,避免链首用全局 retry 的不一致 - retry = route.retry if route else cfg.server.retry - # 配置了 route.timeout 则整条链(含链首)统一用之,否则用 server.timeout - timeout = route.timeout if (route and route.timeout) else cfg.server.timeout - adapter = pkey = pconfig = None + # P0-1: 熔断检查 + cb = _circuits.get(pkey) + if not cb.allow(): + log.warning("[rid=%s] 供应商 %s 熔断中,拒绝 [%s]", rid, pkey, model_id) + raise HTTPException(503, f"供应商 {pconfig.name} 熔断中,请稍后重试") + + # 供应商级/全局级限流(Bug 1;路由级限流随路由功能一并移除) + try: + _enforce_rate_limit(cfg, request, pkey, model_id) + except HTTPException: + # 第六轮: 刚才是 cb.allow() 放行的(可能消耗了 HALF_OPEN 试探令牌), + # 但请求根本没发出去,必须把令牌还回去,否则该供应商会被半开状态永久拒绝。 + cb.release_probe() + raise + + ok = False + result = None + last_err = None + t0 = time.time() + retry = max(1, cfg.server.retry) + for attempt in range(1, retry + 1): try: - adapter, pkey, pconfig = _get_adapter(model_id, timeout, cfg) - except ModelNotFoundError as e: - errors.append({"model": model_id, "error": str(e)}) - log.warning("[rid=%s] fallback 跳过 [%s]: %s", rid, model_id, e) - continue - except AuthConfigError as e: - errors.append({"model": model_id, "error": str(e)}) - continue + result = await adapter.chat( + model=model_id, messages=messages, + temperature=req.temperature, top_p=req.top_p, + max_tokens=req.max_tokens, stop=req.stop, + tools=tools, tool_choice=tool_choice, + extra=req.upstream_extra() or None, + session_id=_upstream_session(req.conversation_id, caller), + ) + ok = True + break + except Exception as e: + last_err = e + log.error("[rid=%s] 上游失败 [%s] attempt=%d/%d: %s", rid, model_id, attempt, retry, e) + if attempt < retry: + await asyncio.sleep(0.3 * attempt) - # P0-1: 熔断检查 - cb = _circuits.get(pkey) - if not cb.allow(): - errors.append({"model": model_id, "error": f"供应商 {pconfig.name} 熔断中", "circuit": "open"}) - log.warning("[rid=%s] 供应商 %s 熔断跳过 [%s]", rid, pkey, model_id) - continue + cb.record(ok) + if not ok: + _metrics.record_call(model_id, pkey, ok=False) + _metrics.record_latency(model_id, time.time() - t0) # Bug 5: 失败也算耗时 + # 第六轮: 键名从 `error` 改为 `message` —— OpenAI 规范要求 error.message, + # 此前客户端拿到的是 {"error":{"error": ...}},SDK 解析不到 message。 + raise HTTPException(502, {"message": f"模型 {model_id} 调用失败", + "errors": [{"model": model_id, "error": str(last_err)}]}) - # 路由级/供应商级限流(Bug 1) - try: - _enforce_rate_limit(cfg, request, route, pkey, model_id) - except HTTPException as e: - # 当前模型所在维度打满 → 切下一个模型(路由内其它模型可继续) - # 第六轮: 刚才是 cb.allow() 放行的(可能消耗了 HALF_OPEN 试探令牌), - # 但请求根本没发出去,必须把令牌还回去,否则该供应商会被半开状态永久拒绝。 - cb.release_probe() - errors.append({"model": model_id, "error": f"限流: {e.detail}", "rate_limited": True}) - log.warning("[rid=%s] [%s] 限流,尝试下一个模型", rid, model_id) - continue - - ok = False - result = None - last_err = None - t0 = time.time() - for attempt in range(1, max(1, retry) + 1): - try: - result = await adapter.chat( - model=model_id, messages=messages, - temperature=req.temperature, top_p=req.top_p, - max_tokens=req.max_tokens, stop=req.stop, - tools=tools, tool_choice=tool_choice, - extra=req.upstream_extra() or None, - session_id=_upstream_session(req.conversation_id, caller), - ) - ok = True - break - except Exception as e: - last_err = e - log.error("[rid=%s] 上游失败 [%s] attempt=%d/%d: %s", rid, model_id, attempt, retry, e) - if attempt < retry: - await asyncio.sleep(0.3 * attempt) - - cb.record(ok) - if not ok: - errors.append({"model": model_id, "error": str(last_err)}) - _metrics.record_call(model_id, pkey, ok=False, route=route.name if route else None) - _metrics.record_latency(model_id, time.time() - t0) # Bug 5: 失败也算耗时 - continue - - # 成功 - usage = result.get("usage") or {} - _metrics.record_call(model_id, pkey, ok=True, route=route.name if route else None) - _metrics.record_tokens(int(usage.get("prompt_tokens", 0)), int(usage.get("completion_tokens", 0))) - _metrics.record_latency(model_id, time.time() - t0) - if model_id != chain[0]: - _metrics.record_fallback(route.name if route else None) - served_model = model_id - _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, - "model": req.model, "served": served_model, "provider": pkey, - "caller": caller, "ok": True, "stream": False, - "response": _logsafe(_resp_text(result), 2000)}) - resp = { - "id": result.get("id", "") or f"chatcmpl-{rid}", "object": "chat.completion", - "created": int(time.time()), "model": req.model, - "system_fingerprint": f"gw-api-{pkey}", - "served_model": served_model, "provider": pkey, "provider_name": pconfig.name, - "choices": result.get("choices", []), "usage": usage, "x_request_id": rid, - } - _attach_cost(resp, served_model, usage, cfg) - if route: - resp["route"] = {"requested": req.model, "served": served_model, - "fallback_used": served_model != chain[0]} - return resp - - # Bug 1 修复: 全部模型失败时不再计入 fallback_total(它是"成功降级"计数), - # 否则 fallback_total 会被纯失败污染。失败已由各次 record_call(ok=False) 记录。 - # 第六轮: 键名从 `error` 改为 `message` —— OpenAI 规范要求 error.message, - # 此前客户端拿到的是 {"error":{"error": ...}},SDK 解析不到 message。 - raise HTTPException(502, {"message": f"路由 {req.model} 全部模型调用失败", "errors": errors}) + usage = result.get("usage") or {} + _metrics.record_call(model_id, pkey, ok=True) + _metrics.record_tokens(int(usage.get("prompt_tokens", 0)), int(usage.get("completion_tokens", 0))) + _metrics.record_latency(model_id, time.time() - t0) + _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, + "model": req.model, "served": model_id, "provider": pkey, + "caller": caller, "ok": True, "stream": False, + "response": _logsafe(_resp_text(result), 2000)}) + resp = { + "id": result.get("id", "") or f"chatcmpl-{rid}", "object": "chat.completion", + "created": int(time.time()), "model": req.model, + "system_fingerprint": f"gw-api-{pkey}", + "served_model": model_id, "provider": pkey, "provider_name": pconfig.name, + "choices": result.get("choices", []), "usage": usage, "x_request_id": rid, + } + _attach_cost(resp, model_id, usage, cfg) + return resp # ── 流式 fallback(元信息 + 熔断 + 统一 [DONE]) ── -def _stream_response(request, req, chain, route, messages, tools): +def _sse_error(rid, caller, req, model_id, detail, provider=None) -> str: + """流式路径的统一错误帧(error 帧 + [DONE] 帧),并按失败记一条访问日志。 + + 第二十七轮: 不再有"这个模型失败就换下一个"的余地(路由/fallback 链已移除), + 所以错误帧在这里一次成型,调用方 yield 一下就收尾。 + """ + _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, + "model": req.model, "served": model_id, "provider": provider, + "caller": caller, "ok": False, "stream": True, "error": str(detail)[:200], + "response": ""}) + err = {"error": detail, "errors": [{"model": model_id, "error": detail}]} + return (f"data: {json.dumps(err, ensure_ascii=False)}\n\n" + f"data: [DONE]\n\n") + + +def _stream_response(request, req, model_id, messages, tools): cfg = _cfg() rid = getattr(request.state, "rid", "-") caller = getattr(request.state, "caller", "anonymous") async def event_stream(): - errors: list[dict] = [] - served = None - # 本轮修复: 日志累加器必须在链循环**外**初始化。此前它写在循环体 try 内, - # 如果链上每个模型都在 _get_adapter / 熔断 / 限流阶段被 continue 掉,那次赋值 - # 永远不会执行 → 全失败分支引用它会 NameError,SSE 变成 500 而不是给客户端错误帧。 + # 本轮修复: 日志累加器必须在 try 外初始化。此前它写在链循环体 try 内, + # 如果模型在 _get_adapter / 熔断 / 限流阶段就被拒,那些赋值永远不会执行 → + # 失败分支引用它会 NameError,SSE 变成 500 而不是给客户端错误帧。 resp_parts: list[str] = [] # 流式输出正文累积(供访问日志) reason_parts: list[str] = [] # 流式推理段累积(正文为空时回落,供访问日志) + timeout = cfg.server.timeout try: - for model_id in chain: - # #2: 配了 route.timeout 整条链(含链首)统一用之 - timeout = route.timeout if (route and route.timeout) else cfg.server.timeout - try: - adapter, pkey, pconfig = _get_adapter(model_id, timeout, cfg) - except (ModelNotFoundError, AuthConfigError) as e: - errors.append({"model": model_id, "error": str(e)}) - continue + try: + adapter, pkey, pconfig = _get_adapter(model_id, timeout, cfg) + except ModelNotFoundError as e: + yield _sse_error(rid, caller, req, model_id, str(e)) + return + except AuthConfigError as e: + yield _sse_error(rid, caller, req, model_id, + f"供应商 {e.provider_key} 未配置可用的 API Key({e.detail})") + return - cb = _circuits.get(pkey) - if not cb.allow(): - errors.append({"model": model_id, "error": f"供应商 {pconfig.name} 熔断中", "circuit": "open"}) - continue + cb = _circuits.get(pkey) + if not cb.allow(): + yield _sse_error(rid, caller, req, model_id, f"供应商 {pconfig.name} 熔断中,请稍后重试", pkey) + return - # 路由级/供应商级限流(Bug 1):流式里用非抛异常方式,打满则跳过该模型 - try: - _enforce_rate_limit(cfg, request, route, pkey, model_id) - except HTTPException as e: - cb.release_probe() # 第六轮: 归还试探令牌(请求没发出去) - errors.append({"model": model_id, "error": f"限流: {e.detail}", "rate_limited": True}) - continue + # 供应商级/全局级限流(Bug 1;路由级限流随路由功能一并移除) + try: + _enforce_rate_limit(cfg, request, pkey, model_id) + except HTTPException as e: + cb.release_probe() # 第六轮: 归还试探令牌(请求没发出去) + yield _sse_error(rid, caller, req, model_id, f"限流: {e.detail}", pkey) + return - t0 = time.time() - stream_gen = None # 第六轮: 显式持有生成器,收尾时 aclose() - try: - committed = False # 是否已向客户端输出过 content(一旦输出就无法回退模型) - emitted_meta = False - stream_usage: dict = {} - resp_parts.clear() # 每个模型单独累积,失败回退时不混入上一模型输出 - reason_parts.clear() - # 第六轮: 用 stream_gen 承接后再 async for,配合下面的 finally aclose()。 - # 上游 chat_stream 内部用 `async with self._lease()` 持有 client 占用, - # 若只 break 不 aclose,生成器要等 GC 才释放 —— 热重载的延迟关闭会 - # 一直等不到 _active==0,连接池迟迟不关。 - stream_gen = adapter.chat_stream( - model=model_id, messages=messages, - temperature=req.temperature, top_p=req.top_p, - max_tokens=req.max_tokens, stop=req.stop, - tools=tools, tool_choice=req.tool_choice, - extra=req.upstream_extra() or None, - session_id=_upstream_session(req.conversation_id, caller), - ) - async for chunk in stream_gen: - if chunk.get("done"): - break - # 功能 8: 累计上游 chunk 的 usage(若有) - if chunk.get("usage"): - u = chunk["usage"] - stream_usage["prompt_tokens"] = stream_usage.get("prompt_tokens", 0) + int(u.get("prompt_tokens", 0)) - stream_usage["completion_tokens"] = stream_usage.get("completion_tokens", 0) + int(u.get("completion_tokens", 0)) - stream_usage["total_tokens"] = stream_usage.get("total_tokens", 0) + int(u.get("total_tokens", 0)) - # Bug 5 修复: meta 延迟到首个 content 到达时才发送 —— - # 若模型在发出 content 前就失败,客户端还没收到任何字节,可安全回退下一个模型。 - if chunk.get("choices"): - if not emitted_meta: - # 第六轮: 网关元信息帧此前不是合法 chunk(缺 id/created/choices, - # object 也写成 chat.completion)→ 严格校验的 SDK 直接报错。 - # 现在补齐为合法的 chat.completion.chunk(choices 为空是 OpenAI - # 自己在 usage-only 帧里的既有约定),自定义字段照旧带上。 - meta = {"id": f"chatcmpl-{rid}", "object": "chat.completion.chunk", - "created": int(time.time()), "model": req.model, - "choices": [], - "served_model": model_id, "provider": pkey, - "provider_name": pconfig.name, "x_request_id": rid} - if route and model_id != chain[0]: - meta["route"] = {"requested": req.model, "served": model_id, "fallback_used": True} - yield f"data: {json.dumps(meta, ensure_ascii=False)}\n\n" - emitted_meta = True - committed = True - served = model_id - # M4: 累积流式输出正文到日志(推理段单独攒,正文为空时才回落) - for _c in chunk.get("choices", []): - _d = _c.get("delta") or {} - if _d.get("content"): - resp_parts.append(_d["content"]) - if _d.get("reasoning_content"): - reason_parts.append(_d["reasoning_content"]) - yield f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n" - cb.record(True) - if served: - _metrics.record_call(served, pkey, ok=True, route=route.name if route else None) - _metrics.record_latency(served, time.time() - t0) - # 本轮修复: 成功降级才计入 fallback_total(此前的流式路径恰好记反了 —— - # 全失败时计数、成功降级时不计,与非流式语义相反)。 - if route and served != chain[0]: - _metrics.record_fallback(route.name) - _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, - "model": req.model, "served": served, "provider": pkey, - "caller": caller, "ok": True, "stream": True, - "response": _logsafe(_stream_log_text(resp_parts, reason_parts), 2000)}) - # 功能 8: 流式末尾发一个 usage 块(累计自上游 chunks 的 usage,若有) - if stream_usage: - # 第六轮: 同上,末尾 usage 帧也补齐 id/created(object 本就是 chunk) - uf = {'id': f'chatcmpl-{rid}', 'object': 'chat.completion.chunk', 'created': int(time.time()), - 'model': req.model, 'choices': [], 'usage': stream_usage, - 'served_model': served, 'provider': pkey} - # M2: 流式 usage 帧同样带上 token 成本估算,前端能显示费用 - _attach_cost(uf, served, stream_usage, cfg) - yield f"data: {json.dumps(uf, ensure_ascii=False)}\n\n" - yield "data: [DONE]\n\n" - return - except Exception as e: - cb.record(False) - _metrics.record_call(model_id, pkey, ok=False, route=route.name if route else None) - _metrics.record_latency(model_id, time.time() - t0) # Bug 5: 失败也算耗时 - if committed: - # 已向客户端输出过 content,无法回退 → 发错误块并结束(避免破坏已输出的 SSE 流) - _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, - "model": req.model, "served": model_id, "provider": pkey, - "caller": caller, "ok": False, "stream": True, "error": str(e)[:200], - "response": _logsafe(_stream_log_text(resp_parts, reason_parts), 2000)}) - yield f"data: {json.dumps({'error': str(e), 'served_model': model_id}, ensure_ascii=False)}\n\n" - yield "data: [DONE]\n\n" - return - # 尚未输出任何 content → 可安全回退到链上下一个模型(Bug 5 修复) - errors.append({"model": model_id, "error": str(e)}) - continue - finally: - # 第六轮: 无论正常收尾 / break / 异常 / 取消,都立刻关掉上游生成器, - # 让 lease 马上归还(不依赖 GC)。已在结束时 aclose 是幂等的。 - if stream_gen is not None: - try: - await stream_gen.aclose() - except Exception: - pass - - # 本轮修复: 全失败**不再**计入 fallback_total(它表示"成功降级"次数)。 - # 失败已由各自的 record_call(ok=False) 记录;此处此前会污染指标(与非流式修复不一致)。 - _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, - "model": req.model, "served": None, "provider": None, - "caller": caller, "ok": False, "stream": True, - "response": _logsafe(_stream_log_text(resp_parts, reason_parts), 2000), "error": "所有模型调用失败"}) - yield f"data: {json.dumps({'error': '所有模型调用失败', 'errors': errors}, ensure_ascii=False)}\n\n" - yield "data: [DONE]\n\n" + t0 = time.time() + stream_gen = None # 第六轮: 显式持有生成器,收尾时 aclose() + try: + emitted_meta = False + stream_usage: dict = {} + # 第六轮: 用 stream_gen 承接后再 async for,配合下面的 finally aclose()。 + # 上游 chat_stream 内部用 `async with self._lease()` 持有 client 占用, + # 若只 break 不 aclose,生成器要等 GC 才释放 —— 热重载的延迟关闭会 + # 一直等不到 _active==0,连接池迟迟不关。 + stream_gen = adapter.chat_stream( + model=model_id, messages=messages, + temperature=req.temperature, top_p=req.top_p, + max_tokens=req.max_tokens, stop=req.stop, + tools=tools, tool_choice=req.tool_choice, + extra=req.upstream_extra() or None, + session_id=_upstream_session(req.conversation_id, caller), + ) + async for chunk in stream_gen: + if chunk.get("done"): + break + # 功能 8: 累计上游 chunk 的 usage(若有) + if chunk.get("usage"): + u = chunk["usage"] + stream_usage["prompt_tokens"] = stream_usage.get("prompt_tokens", 0) + int(u.get("prompt_tokens", 0)) + stream_usage["completion_tokens"] = stream_usage.get("completion_tokens", 0) + int(u.get("completion_tokens", 0)) + stream_usage["total_tokens"] = stream_usage.get("total_tokens", 0) + int(u.get("total_tokens", 0)) + # Bug 5 修复: meta 帧在首个 content 到达时才发送 —— 第六轮补齐为合法 chunk + # (缺 id/created/choices 的"元信息帧"会被严格校验的 SDK 直接拒掉)。 + if chunk.get("choices") and not emitted_meta: + meta = {"id": f"chatcmpl-{rid}", "object": "chat.completion.chunk", + "created": int(time.time()), "model": req.model, + "choices": [], + "served_model": model_id, "provider": pkey, + "provider_name": pconfig.name, "x_request_id": rid} + yield f"data: {json.dumps(meta, ensure_ascii=False)}\n\n" + emitted_meta = True + # M4: 累积流式输出正文到日志(推理段单独攒,正文为空时才回落) + for _c in chunk.get("choices", []): + _d = _c.get("delta") or {} + if _d.get("content"): + resp_parts.append(_d["content"]) + if _d.get("reasoning_content"): + reason_parts.append(_d["reasoning_content"]) + yield f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n" + cb.record(True) + _metrics.record_call(model_id, pkey, ok=True) + _metrics.record_latency(model_id, time.time() - t0) + _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, + "model": req.model, "served": model_id, "provider": pkey, + "caller": caller, "ok": True, "stream": True, + "response": _logsafe(_stream_log_text(resp_parts, reason_parts), 2000)}) + # 功能 8: 流式末尾发一个 usage 块(累计自上游 chunks 的 usage,若有) + if stream_usage: + uf = {'id': f'chatcmpl-{rid}', 'object': 'chat.completion.chunk', 'created': int(time.time()), + 'model': req.model, 'choices': [], 'usage': stream_usage, + 'served_model': model_id, 'provider': pkey} + # M2: 流式 usage 帧同样带上 token 成本估算,前端能显示费用 + _attach_cost(uf, model_id, stream_usage, cfg) + yield f"data: {json.dumps(uf, ensure_ascii=False)}\n\n" + yield "data: [DONE]\n\n" + return + except Exception as e: + cb.record(False) + _metrics.record_call(model_id, pkey, ok=False) + _metrics.record_latency(model_id, time.time() - t0) # Bug 5: 失败也算耗时 + _access.add({"method": "POST", "kind": "chat", "path": "/v1/chat/completions", "rid": rid, + "model": req.model, "served": model_id, "provider": pkey, + "caller": caller, "ok": False, "stream": True, "error": str(e)[:200], + "response": _logsafe(_stream_log_text(resp_parts, reason_parts), 2000)}) + # 第六轮: 已向客户端输出过 content 时本来就无法回退(第二十七轮起更没有下一个 + # 模型可退),统一发错误帧 + [DONE] 收尾,避免破坏已输出的 SSE 流。 + yield f"data: {json.dumps({'error': str(e), 'served_model': model_id}, ensure_ascii=False)}\n\n" + yield "data: [DONE]\n\n" + return + finally: + # 第六轮: 无论正常收尾 / break / 异常 / 取消,都立刻关掉上游生成器, + # 让 lease 马上归还(不依赖 GC)。已在结束时 aclose 是幂等的。 + if stream_gen is not None: + try: + await stream_gen.aclose() + except Exception: + pass finally: # 第六轮: in_flight 已改为在 middleware 的 finally 中统一递减(见 in_flight_tracker), # 此处不再递减 —— 此前流式请求会被双重递减(详见那边的说明)。 @@ -1616,18 +1863,13 @@ def _stream_response(request, req, chain, route, messages, tools): async def chat_batch(request: Request, req: BatchRequest, _auth=Depends(require_auth)): cfg = _cfg() rid = getattr(request.state, "rid", "-") - # Bug 15: 批量入口同样对未知模型/路由直接 400 - if cfg.find_route(req.model) is None and cfg.find_provider_for_model(req.model) is None: - raise HTTPException(400, f"未知模型或路由: {req.model}") + # Bug 15: 批量入口同样对未知模型直接 400(第二十七轮: 这条入口也是 chat/completions 线) + model_id = _require_protocol_model(cfg, req.model, PROTOCOL_CHAT) # 本轮修复: 批量条数上限(此前只受请求体 2MB 限制,可一次塞进上千条 → 自伤型 DoS) if not req.items: raise HTTPException(400, "items 不能为空") if len(req.items) > cfg.server.batch_max_items: raise HTTPException(400, f"items 数量 {len(req.items)} 超过上限 {cfg.server.batch_max_items}") - # 本轮优化: 模型链只解析一次(此前每个 item 都在并发协程里重算一遍) - chain, route = _route_chain(cfg, req.model) - if not chain: - raise HTTPException(400, f"路由 {req.model} 未配置有效模型链") # 本轮修复: 并发上限钳制(此前完全由客户端 max_concurrency 决定,可传 1e5 打爆上游/自身) limit = min(max(1, req.max_concurrency), cfg.server.batch_max_concurrency) sem = asyncio.Semaphore(limit) @@ -1644,7 +1886,7 @@ async def chat_batch(request: Request, req: BatchRequest, _auth=Depends(require_ max_tokens=item.max_tokens, stream=False, extra=passthrough, ) - r = await _chat_fallback(cfg, rid, "batch", request, sub, chain, route, messages, None) + r = await _chat_once(cfg, rid, "batch", request, sub, model_id, messages, None) return {"index": i, "label": item.label, "ok": True, "result": r} except HTTPException as e: # 第六轮: detail 是结构化 dict(如全部模型失败)时,str() 会产出 Python repr, @@ -1674,10 +1916,9 @@ async def images_generations(request: Request, req: ImageRequest, _auth=Depends( cfg = _cfg() rid = getattr(request.state, "rid", "-") caller = getattr(request.state, "caller", "anonymous") - # Bug 4: 支持路由别名(如 wjd)→ 解析到第一个 type==image 的模型 resolved = _resolve_media_model(cfg, req.model, "image") if resolved is None: - raise HTTPException(404, f"未知图像模型/路由: {req.model}(没有可用的 image 类型模型)") + raise HTTPException(404, f"未知图像模型: {req.model}(没有可用的 image 类型模型)") pkey, pconfig, minfo, token, real_model = resolved # 顺序修复(本轮):先做能力检查(501)、再限流 —— 否则"不支持图像的供应商"也会白扣一个限流 token。 if not pconfig.media.supports_image: @@ -1921,10 +2162,9 @@ async def videos_create(request: Request, req: VideoCreateRequestExtended, _auth rid = getattr(request.state, "rid", "-") caller = getattr(request.state, "caller", "anonymous") _validate_video_request(req) # 第十一轮 #3:入站按 Agnes 契约校验 mode/媒体/时长 - # Bug 4: 支持路由别名(如 wjd)→ 解析到第一个 type==video 的模型 resolved = _resolve_media_model(cfg, req.model, "video") if resolved is None: - raise HTTPException(404, f"未知视频模型/路由: {req.model}(没有可用的 video 类型模型)") + raise HTTPException(404, f"未知视频模型: {req.model}(没有可用的 video 类型模型)") pkey, pconfig, minfo, token, real_model = resolved # 顺序修复(本轮):先能力检查(501),再校验回调地址(400),最后限流 —— 避免无效请求白扣 token。 if not pconfig.media.supports_video: @@ -2055,12 +2295,12 @@ async def videos_poll(request: Request, video_id: str, model: str = "", wait: bo # 本轮重写:供应商定位不再"盲猜"。 # 视频任务 ID 是**供应商私有**的,查错供应商必然 404/502,且静默误导。 - # 策略:给了 model/路由别名 → 用与创建端点相同的解析逻辑(_resolve_media_model); + # 策略:给了 model → 用与创建端点相同的解析逻辑(_resolve_media_model); # 没给 model → 仅当全站只有一家支持视频的供应商时才推断,否则要求显式指定。 if model: resolved = _resolve_media_model(cfg, model, "video") if resolved is None: - raise HTTPException(404, f"未知视频模型/路由: {model}(没有可用的 video 类型模型)") + raise HTTPException(404, f"未知视频模型: {model}(没有可用的 video 类型模型)") pkey, pconfig, _minfo, _token, pmodel_id = resolved else: video_providers = [k for k, v in cfg.providers.items() if v.media.supports_video] @@ -2422,7 +2662,8 @@ async def metrics_prometheus(request: Request, _auth=Depends(require_auth)): lines = [ "# API Gateway Prometheus 指标", f'api_gateway_uptime_seconds {snap["uptime_seconds"]}', - f'api_gateway_fallback_total {snap["fallback_total"]}', + # 第二十七轮: api_gateway_fallback_total 随「路由别名 / fallback 链」一起删除 —— + # 快照里已没有 fallback_total 字段,这里再读就是 KeyError(曾被 live_check 抓到 500)。 f'api_gateway_tokens_prompt_total {snap["tokens"]["prompt"]}', f'api_gateway_tokens_completion_total {snap["tokens"]["completion"]}', # 功能 11: 在途请求 gauge @@ -2445,11 +2686,12 @@ async def metrics_prometheus(request: Request, _auth=Depends(require_auth)): lines.append(f'api_gateway_circuit_state{{provider="{pl}"}} {state_val}') lines.append(f'api_gateway_circuit_recent_failures{{provider="{pl}"}} {c["recent_failures"]}') # 第六轮: 补 HELP/TYPE(promtool / Grafana 的规范要求;此前只有裸指标行) + # 第二十七轮: fallback_total 的 HELP/TYPE 已随路由/fallback 链删除(连同上面的指标行)。 + # 注意别把"已删除"写成裸注释行塞进导出的正文里:正文里只放真正存在的指标的 + # HELP/TYPE,说明留在代码注释里(测试会校验声明与导出的一一对应)。 help_lines = [ "# HELP api_gateway_uptime_seconds 网关运行时长(秒)", "# TYPE api_gateway_uptime_seconds gauge", - "# HELP api_gateway_fallback_total 成功降级到备选模型的次数(纯失败不计)", - "# TYPE api_gateway_fallback_total counter", "# HELP api_gateway_tokens_prompt_total 上游 prompt token 累计", "# TYPE api_gateway_tokens_prompt_total counter", "# HELP api_gateway_tokens_completion_total 上游 completion token 累计", @@ -2514,7 +2756,7 @@ async def logs(request: Request, n: int = 100, _auth=Depends(require_auth)): # ───────────────────────────────────── @app.get("/", include_in_schema=False) async def root(request: Request, _auth=Depends(require_auth_or_open)): - # 第六轮: / 也纳入 health_auth 保护(会列出供应商/路由/端点清单);_auth 未使用。 + # 第六轮: / 也纳入 health_auth 保护(会列出供应商/端点清单);_auth 未使用。 if "text/html" in (request.headers.get("accept") or ""): # 浏览器访问根路径 → 前端登录页;API 调用(Accept 不含 text/html)仍返回 JSON 导航 return RedirectResponse("/web/") @@ -2523,7 +2765,11 @@ async def root(request: Request, _auth=Depends(require_auth_or_open)): "name": "API Gateway", "version": VERSION, "endpoints": { - "chat": "POST /v1/chat/completions(支持路由别名 wjd / 流式 / fallback / 熔断)", + "chat": "POST /v1/chat/completions(OpenAI Chat Completions 线)", + # 第二十七轮「三线直传」: 入站是什么格式,出站就是什么格式 —— 网关不翻译协议, + # 只在模型声明支持该线时原样转发(不支持则 400 说清它支持哪几条线)。 + "messages": "POST /v1/messages(Anthropic Messages 线,鉴权也认 x-api-key)", + "responses": "POST /v1/responses(OpenAI Responses 线)", "batch": "POST /v1/chat/completions/batch(并发多 prompt)", "images": "POST /v1/images/generations", "videos_create": "POST /v1/videos(异步,?wait 或 callback_url)", @@ -2535,7 +2781,6 @@ async def root(request: Request, _auth=Depends(require_auth_or_open)): "docs": "GET /docs(Swagger UI)", }, "providers": list(cfg.providers.keys()), - "routes": list(cfg.routes.keys()), } @@ -3074,9 +3319,9 @@ async def config_get(request: Request, _auth=Depends(require_auth)): - `config`:面板表单用的值(结构与 gateway.yaml 一一对应;密钥只回报"配没配") - `yaml`:**文件当前原文**(页面上可展开看,不做二次加工 —— 你看到的就是盘上的) - - `warnings`:配置体检结论(通配 CORS、缺 Key、路由引用了不存在的模型 …) + - `warnings`:配置体检结论(通配 CORS、缺 Key 等) - `restart_keys`:改了必须重启进程才生效的键(页面按此在输入框旁标注) - - `model_ids` / `route_names`:给路由、定价的下拉用(不用前端再猜) + - `model_ids`:给模型、定价的下拉用(不用前端再猜) """ _require_admin(request) cfg = _cfg() @@ -3093,9 +3338,7 @@ async def config_get(request: Request, _auth=Depends(require_auth)): "path": str(CONFIG_PATH), "restart_keys": list(gw_settings.RESTART_KEYS), "warnings": validate_warnings(cfg), - "route_warnings": cfg.validate_routes(), "model_ids": [m["id"] for m in cfg.all_models()], - "route_names": list(cfg.routes.keys()), "model_types": list(gw_settings.MODEL_TYPES), "media_keys": list(gw_settings.MEDIA_KEYS), "mtime": _config_mtime(), @@ -3138,8 +3381,8 @@ async def config_put(request: Request, _auth=Depends(require_auth)): restart = gw_settings.restart_keys_changed(old_cfg, new_cfg) await _apply_config(new_cfg, source=f"页面保存YAML({request.state.account})") _state["mtime"] = _config_mtime() - log.warning("配置已由 %s 通过管理页面以 YAML 原文保存并生效(%d 供应商 / %d 路由)", - request.state.account, len(new_cfg.providers), len(new_cfg.routes)) + log.warning("配置已由 %s 通过管理页面以 YAML 原文保存并生效(%d 供应商)", + request.state.account, len(new_cfg.providers)) return {"status": "ok", "applied": True, "restart_required": restart, "warnings": warns, "path": info.get("path", str(CONFIG_PATH)), "backup": info.get("backup", ""), "yaml": text, "mtime": _state["mtime"]} @@ -3158,8 +3401,8 @@ async def config_put(request: Request, _auth=Depends(require_auth)): restart = gw_settings.restart_keys_changed(old_cfg, new_cfg) await _apply_config(new_cfg, source=f"页面保存({request.state.account})") _state["mtime"] = _config_mtime() - log.warning("配置已由 %s 通过管理页面保存并生效(%d 供应商 / %d 路由 / %d 字节)", - request.state.account, len(new_cfg.providers), len(new_cfg.routes), info.get("bytes", 0)) + log.warning("配置已由 %s 通过管理页面保存并生效(%d 供应商 / %d 字节)", + request.state.account, len(new_cfg.providers), info.get("bytes", 0)) return { "status": "ok", "applied": True, @@ -3199,7 +3442,7 @@ async def health(request: Request, _auth=Depends(require_auth_or_open)): "uptime_seconds": snap.get("uptime_seconds"), "in_flight": snap.get("in_flight", 0), "in_flight_peak": snap.get("in_flight_peak", 0), - "fallback_total": snap.get("fallback_total", 0), + # 第二十七轮: fallback_total 已随路由/fallback 链删除(这里原本恒为 0 的死字段) "tokens": snap.get("tokens", {}), "providers": { pkey: { @@ -3213,7 +3456,6 @@ async def health(request: Request, _auth=Depends(require_auth_or_open)): } for pkey, pconfig in cfg.providers.items() }, - "routes": list(cfg.routes.keys()), "models": model_health, } diff --git a/app/metrics.py b/app/metrics.py index 57b581b..26002d9 100644 --- a/app/metrics.py +++ b/app/metrics.py @@ -56,15 +56,13 @@ _BODY_MAX = 2000 # ── 用量指标 ────────────────────────────────────────── class Metrics: - """按 模型/供应商/路由 维度统计调用次数、fallback、token、耗时。""" + """按 模型/供应商 维度统计调用次数、token、耗时。""" def __init__(self, enabled: bool = True, started_at: float | None = None): self.enabled = enabled self._lock = threading.Lock() self.calls: dict[str, dict[str, int]] = {} # model -> {ok, fail} self.providers: dict[str, dict[str, int]] = {} # provider -> {ok, fail} - self.routes: dict[str, dict[str, int]] = {} # route -> {ok, fallback} - self.fallback_total: int = 0 self.tokens_prompt: int = 0 self.tokens_completion: int = 0 self.latency_samples: dict[str, deque[float]] = {} # model -> 最近 N 次耗时 @@ -72,7 +70,7 @@ class Metrics: self.in_flight: int = 0 self.in_flight_peak: int = 0 # M5+: 按账号维度的聚合(普通账号只看到自己的调用;admin 看全局)。 - # 结构:account -> {"calls","providers","routes","tokens","latency","fallback"} + # 结构:account -> {"calls","providers","tokens","latency"} self.by_account: dict[str, dict[str, Any]] = {} # Bug 4: 热重载时保留 started_at,uptime 不因重载清零 self.started_at: float = started_at if started_at is not None else time.time() @@ -82,9 +80,9 @@ class Metrics: acc = self.by_account.get(account) if acc is None: acc = { - "calls": {}, "providers": {}, "routes": {}, + "calls": {}, "providers": {}, "tokens": {"prompt": 0, "completion": 0}, - "latency": {}, "fallback": 0, + "latency": {}, } self.by_account[account] = acc return acc @@ -109,11 +107,10 @@ class Metrics: def _bucket(d: dict, k: str) -> dict[str, int]: return d.setdefault(k, {"ok": 0, "fail": 0}) - def record_call(self, model: str, provider: str | None, ok: bool, route: str | None = None, + def record_call(self, model: str, provider: str | None, ok: bool, account: str | None = None) -> None: """记录一次调用结果。 - route 维度(Bug 8 修复):传入 route 时同时累加 routes 的 ok/fail。 account(M5+):未显式传入时取请求级 ACCOUNT 上下文;非 anonymous 时额外累加 到该账号自己的聚合桶,供「只看本人的指标」使用。 """ @@ -124,31 +121,12 @@ class Metrics: self._bucket(self.calls, model)["ok" if ok else "fail"] += 1 if provider: self._bucket(self.providers, provider)["ok" if ok else "fail"] += 1 - if route: - self._bucket(self.routes, route)["ok" if ok else "fail"] += 1 if a and a != "anonymous": acc = self._acc(a) self._bucket(acc["calls"], model)["ok" if ok else "fail"] += 1 if provider: self._bucket(acc["providers"], provider)["ok" if ok else "fail"] += 1 - if route: - self._bucket(acc["routes"], route)["ok" if ok else "fail"] += 1 - def record_fallback(self, route: str | None, account: str | None = None) -> None: - if not self.enabled: - return - a = account or ACCOUNT.get() - with self._lock: - self.fallback_total += 1 - if route: - self._bucket(self.routes, route).setdefault("fallback", 0) - self.routes[route]["fallback"] += 1 - if a and a != "anonymous": - acc = self._acc(a) - acc["fallback"] += 1 - if route: - b = self._bucket(acc["routes"], route) - b["fallback"] = b.get("fallback", 0) + 1 def record_tokens(self, prompt: int, completion: int, account: str | None = None) -> None: if not self.enabled: @@ -198,7 +176,7 @@ class Metrics: if not acc: return dict(uptime_seconds=round(time.time() - self.started_at, 1), in_flight=self.in_flight, in_flight_peak=self.in_flight_peak, - calls={}, providers={}, routes={}, fallback_total=0, + calls={}, providers={}, tokens={"prompt": 0, "completion": 0}, latency={}) lat = {} for m, dq in acc["latency"].items(): @@ -210,8 +188,6 @@ class Metrics: "in_flight_peak": self.in_flight_peak, "calls": {k: dict(v) for k, v in acc["calls"].items()}, "providers": {k: dict(v) for k, v in acc["providers"].items()}, - "routes": {k: dict(v) for k, v in acc["routes"].items()}, - "fallback_total": acc["fallback"], "tokens": dict(acc["tokens"]), "latency": lat, } @@ -231,8 +207,6 @@ class Metrics: "in_flight_peak": self.in_flight_peak, "calls": {k: dict(v) for k, v in self.calls.items()}, "providers": {k: dict(v) for k, v in self.providers.items()}, - "routes": {k: dict(v) for k, v in self.routes.items()}, - "fallback_total": self.fallback_total, "tokens": {"prompt": self.tokens_prompt, "completion": self.tokens_completion}, "latency": latency_out, } diff --git a/app/models.py b/app/models.py index bd534a1..08e0478 100644 --- a/app/models.py +++ b/app/models.py @@ -86,6 +86,37 @@ class ToolDef(BaseModel): CONVERSATION_ID_MAX = 128 +def normalize_conversation_id(v: Any) -> str | None: + """会话 id 的入口校验(第二十六轮,第二十七轮抽成函数给直传端点多路复用)。 + + 它最终会作为 `x-opencode-session` **HTTP header** 发往上游,所以必须在入站拦掉 + 两类值——否则失败发生在组 header 的那一刻,会被误报成"上游挂了": + + ① **非 ASCII**(中文、全角符号等):httpx 组 header 抛 `UnicodeEncodeError` + (实测 `'ascii' codec can't encode characters`),网关把它当成上游失败 + **重试 3 次**(3 次必然都失败)→ 最终回 502「路由 x 全部模型调用失败」+ + 记一次熔断。真实原因是客户端传错了值,却被伪装成上游故障,排查成本极高。 + ② **过长**:header 超限同样会失败,且一行 header 塞几百 KB 毫无意义。 + + 空白串按"未传"处理(返回 None → 上游每次随机会话),与旧行为保持一致。 + 校验失败抛 ValueError,由调用方转成 400(pydantic 校验器与直传端点共用同一份规则)。 + """ + if v is None: + return None + if not isinstance(v, str): + raise ValueError(f"类型必须是字符串,得到 {type(v).__name__}") + v = v.strip() + if not v: + return None + if len(v) > CONVERSATION_ID_MAX: + raise ValueError(f"过长:最多 {CONVERSATION_ID_MAX} 个字符(收到 {len(v)} 个)") + try: + v.encode("ascii") + except UnicodeEncodeError: + raise ValueError("只能包含 ASCII 字符(不能有中文/全角符号/emoji);建议用 uuid4().hex") + return v + + class ChatRequest(BaseModel): # 兼容性修复:OpenAI 官方 SDK 会带不少可选参数(n / presence_penalty / response_format …), # 之前未声明的字段被 pydantic 静默丢弃,导致"网关号称 OpenAI 全兼容但参数悄悄丢失"。 @@ -155,33 +186,8 @@ class ChatRequest(BaseModel): @field_validator("conversation_id") @classmethod def _validate_conversation_id(cls, v: str | None) -> str | None: - """会话 id 的入口校验(第二十六轮)。 - - 它最终会作为 `x-opencode-session` **HTTP header** 发往上游,所以必须在入站拦掉 - 两类值——否则失败发生在组 header 的那一刻,会被误报成"上游挂了": - - ① **非 ASCII**(中文、全角符号等):httpx 组 header 抛 `UnicodeEncodeError` - (实测 `'ascii' codec can't encode characters`),网关把它当成上游失败 - **重试 3 次**(3 次必然都失败)→ 最终回 502「路由 x 全部模型调用失败」+ - 记一次熔断。真实原因是客户端传错了值,却被伪装成上游故障,排查成本极高。 - ② **过长**:header 超限同样会失败,且一行 header 塞几百 KB 毫无意义。 - - 空白串按"未传"处理(返回 None → 上游每次随机会话),与旧行为保持一致。 - """ - if v is None: - return None - v = v.strip() - if not v: - return None - if len(v) > CONVERSATION_ID_MAX: - raise ValueError( - f"过长:最多 {CONVERSATION_ID_MAX} 个字符(收到 {len(v)} 个)") - try: - v.encode("ascii") - except UnicodeEncodeError: - raise ValueError( - "只能包含 ASCII 字符(不能有中文/全角符号/emoji);建议用 uuid4().hex") - return v + """见 normalize_conversation_id(第二十七轮把规则抽成公共函数,直传端点多路复用)。""" + return normalize_conversation_id(v) # ── 图像生成 ───────────────────────────────────────── diff --git a/app/settings.py b/app/settings.py index ec6ff3c..f10a84c 100644 --- a/app/settings.py +++ b/app/settings.py @@ -31,6 +31,8 @@ import yaml from app.config import ( ConfigError, GatewayConfig, + PROTOCOLS, + _default_endpoints, parse_config, validate_config, validate_warnings, @@ -88,11 +90,23 @@ def config_to_view(cfg: GatewayConfig) -> dict[str, Any]: "rate_limit_rps": p.rate_limit_rps, "requires_session_header": bool(p.requires_session_header), "media": {k: getattr(p.media, k) for k in MEDIA_KEYS}, + # 第二十七轮(三线直传): 每条协议线的出站路径/鉴权/固定头。 + # 面板不提供编辑器(在 YAML 全文编辑器里改),但**必须一路带着走**: + # view_to_raw 是按固定键集重建 providers 的,少带这一段 = 管理员在页面上 + # 随手保存一次,OpenCode 的三条线配置就被静默清掉(模型全变"不支持该线")。 + "endpoints": { + name: {"path": ep.path, "auth": ep.auth, "headers": dict(ep.headers or {})} + for name, ep in (p.endpoints or {}).items() + }, "models": [ {"id": m.id, "display_name": m.display_name, "type": m.type, "vision": bool(m.vision), # 第二十五轮: 回传模型级强制参数。面板不提供编辑入口(在 YAML 全文编辑器里改), # 但必须一路带着走 —— 否则读回来再提交一次就会把它弄丢。 - "params": dict(m.params or {})} + "params": dict(m.params or {}), + # 第二十七轮: 模型能走哪几条协议线。这一项**面板可编辑** + # (「供应商与模型」页给每个模型打格式标签),所以两处都要认: + # 读回给页面显示、写盘时以页面提交的为准。 + "protocols": list(m.protocols)} for m in p.models ], } @@ -119,15 +133,6 @@ def config_to_view(cfg: GatewayConfig) -> dict[str, Any]: "auth_fail_limit_per_min": s.auth_fail_limit_per_min, }, "auth": {"session_ttl_minutes": cfg.auth.session_ttl_minutes}, - "routes": { - rname: { - "models": list(r.models or []), - "timeout": r.timeout, - "retry": r.retry, - "rate_limit_rps": r.rate_limit_rps, - } - for rname, r in cfg.routes.items() - }, "circuit": { "enabled": cfg.circuit.enabled, "failure_threshold": cfg.circuit.failure_threshold, @@ -153,6 +158,37 @@ def config_to_view(cfg: GatewayConfig) -> dict[str, Any]: # ── 写入:面板 JSON → 完整 yaml 字典 ─────────────────────── +def _parse_protocols_view(v: Any, where: str) -> list[str]: + """面板提交的模型协议线(第二十七轮三线直传)→ 写盘用的标准名列表。 + + 页面上的格式标签就是这几个标准名(chat/completions | messages | responses), + 写错名字在写盘前就报错(不能等运行时才发现"这个模型永远 400")。 + 列表去重保序;空列表报错(一个模型一条线都不支持多半是手滑)。 + """ + if isinstance(v, str): + raw_items: list[Any] = [v] + elif isinstance(v, (list, tuple)): + raw_items = list(v) + else: + raise ConfigError(f"{where} 必须是列表(可选:{'/'.join(PROTOCOLS)})") + out: list[str] = [] + for item in raw_items: + name = str(item or "").strip() + if name not in PROTOCOLS: + raise ConfigError(f"{where} 里的 {name!r} 不是标准协议名(可选:{'/'.join(PROTOCOLS)})") + if name not in out: + out.append(name) + if not out: + raise ConfigError(f"{where} 不能为空 —— 每个模型至少要声明一条协议线") + return out + + +def _default_endpoints_view() -> dict[str, Any]: + """供应商没在文件里写 endpoints 时,面板/写盘要用的默认三条线。""" + return {name: {"path": ep.path, "auth": ep.auth, "headers": dict(ep.headers or {})} + for name, ep in _default_endpoints().items()} + + def _as_dict(v: Any, field_name: str) -> dict[str, Any]: if v is None: return {} @@ -242,6 +278,11 @@ def view_to_raw(cfg: GatewayConfig, raw: dict[str, Any], view: dict[str, Any]) - "type": str(m.get("type") or "text").strip(), "vision": bool(m.get("vision", False)), }) + # 第二十七轮: 协议线由面板管理(每个模型的格式标签),提交了就以此为准; + # 没提交则沿用文件原文里的值(避免旧前端/脚本提交时把 protocols 清空)。 + if m.get("protocols"): + item["protocols"] = _parse_protocols_view( + m.get("protocols"), f"providers.{key}.models[{i}].protocols") models.append(item) providers[key] = { "name": str(pv.get("name") or key).strip(), @@ -253,6 +294,10 @@ def view_to_raw(cfg: GatewayConfig, raw: dict[str, Any], view: dict[str, Any]) - "extra_headers": _str_map(pv.get("extra_headers"), f"providers.{key}.extra_headers"), "rate_limit_rps": pv.get("rate_limit_rps"), "requires_session_header": bool(pv.get("requires_session_header", False)), + # 第二十七轮: 协议线的出站配置(路径/鉴权/固定头)。 + # 仍然以**文件原文为底**:面板没有这段的编辑器,而它一旦丢失, + # 声明了 messages/responses 的模型会直接变成"供应商没配这条线"。 + "endpoints": base_p.get("endpoints", _default_endpoints_view()), # media 留空 = 该供应商不支持对应媒体(写回时保留原文里的空 {} 语义) "media": { k: (media_in.get(k) if k in media_in else base_media.get(k)) @@ -277,23 +322,6 @@ def view_to_raw(cfg: GatewayConfig, raw: dict[str, Any], view: dict[str, Any]) - merged["session_ttl_minutes"] = av["session_ttl_minutes"] out["auth"] = merged - # ── routes(整段以面板为准:删掉的路由别名就是删掉)── - if "routes" in view: - rv = _as_dict(view.get("routes"), "routes") - routes: dict[str, Any] = {} - for rname, rv1 in rv.items(): - name = str(rname).strip() - if not name: - raise ConfigError("路由别名的名字不能为空") - rv1 = _as_dict(rv1, f"routes.{name}") - routes[name] = { - "models": _str_list(rv1.get("models"), f"routes.{name}.models"), - "timeout": rv1.get("timeout"), - "retry": rv1.get("retry"), - "rate_limit_rps": rv1.get("rate_limit_rps"), - } - out["routes"] = routes - # ── circuit / observability ── for sec, keys in (("circuit", _CIRCUIT_KEYS), ("observability", _OBS_KEYS)): if sec in view: @@ -393,12 +421,6 @@ _SECTION_COMMENTS = { # 账号与 API Key 由页面管理,存在 data/users.json(见 app/users.py): # 管理员在「用户管理」建账号,各账号在「设置 → 我的 API Key」自助铸 Key。 # 本文件里没有任何用户/密钥,只有会话时长这一个开关。 -""", - "routes": """\ -# ── 路由别名 + 按优先级 fallback ──────────────────────── -# model 填别名(如 wjd)时按 models 数组顺序逐个尝试:第 1 个失败/超时/重试耗尽 → 第 2 个 … -# 全失败才 502。timeout = 单模型超时秒(null 用 server.timeout),retry = 切下一个前该模型重试次数。 -# 数组顺序即优先级(第 0 个最高)。可用模型 = 上面 providers 里定义的全部模型 id。 """, "circuit": """\ # ── 供应商熔断 ────────────────────────────────────────── @@ -424,7 +446,7 @@ _SECTION_COMMENTS = { """, } -_SECTION_ORDER = ("providers", "server", "circuit", "observability", "pricing", "download", "auth", "routes") +_SECTION_ORDER = ("providers", "server", "circuit", "observability", "pricing", "download", "auth") def _dump(obj: Any) -> str: diff --git a/data/gateway.yaml b/data/gateway.yaml index 94a79d0..2f01946 100644 --- a/data/gateway.yaml +++ b/data/gateway.yaml @@ -25,7 +25,7 @@ providers: display_name: "Agnes 3.0 Flash" type: "text" # 第十三轮: Agnes 3.0 Flash 文档「输入模态:文本、图像 URL」→ 支持视觉(对话带图)。 - # 实测 data:image/png;base64,... 也能识别。带图请求走路由别名时会被排到链首。 + # 实测 data:image/png;base64,... 也能识别。 vision: true - id: "agnes-2.5-flash" display_name: "Agnes 2.5 Flash" @@ -57,35 +57,71 @@ providers: base_url: "https://opencode.ai/zen/go/v1" auth: token_env: "OPENCODE_API_KEY" - # OpenCode Go 是每月订阅编程模型入口。官方硬性要求每个 /chat/completions - # 带 x-opencode-session(缺失直接 400 MissingSessionID),并建议用专属 UA + # OpenCode Go 是每月订阅编程模型入口。官方硬性要求**三条线**都带 + # x-opencode-session(缺失直接 400 MissingSessionID),并建议用专属 UA # 而非通用库名(实测 UA 不敏感,会话头才是必需)。适配器据此开关在 - # 每次请求动态注入 UUID 会话头(不能放静态 extra_headers——会被 _get_adapter 缓存冻结)。 + # 每次请求动态注入会话头(不能放静态 extra_headers——会被 _get_adapter 缓存冻结)。 requires_session_header: true extra_headers: User-Agent: "api-gateway/2.4.0 (jakon opencode-go)" # 无媒体能力(纯文本编程模型;图像/视频不走本供应商) media: {} + # ── 协议线(第二十七轮「三线直传」)────────────────────────── + # 同一家供应商的三条线共用 base_url,只有「路径 + 鉴权方式 + 固定头」不同。 + # 实测(2026-09-18,真实 key): + # /chat/completions 与 /responses → Authorization: Bearer + # /messages → x-api-key(用 Bearer 打会 401 Missing API key) + # 缺少 anthropic-version 时 /messages 也会被拒 → 作为该线固定头带上。 + # 语义:**这里没写的线 = 该供应商没有这条线**(模型就算声明了也调用不了, + # validate_config 会直接报错拦住)。老板决策:先只做 OpenCode 一家, + # 其余供应商(Agnes / 火山)暂不声明 → 走默认值 = 只认 chat/completions, + # 行为与改造前完全一致。 + endpoints: + chat/completions: + path: "/chat/completions" + auth: "bearer" + messages: + path: "/messages" + auth: "x-api-key" + headers: + anthropic-version: "2023-06-01" + responses: + path: "/responses" + auth: "bearer" + # ── 模型 ───────────────────────────────────────────────── + # models[].protocols = 该模型**能用哪几条线**(入站格式必须与出站一致)。 + # 取值:chat/completions | messages | responses;不写 = 只认 chat/completions。 + # 实测口径(2026-09-18 全量扫描 29 个文档模型 × 3 条线,非文档线另做 + # 「流式 + 工具调用」复核,全部 200):官方文档标注的线**一定**可用; + # 另外有一批模型实测能走多条线(OpenCode 侧自带跨格式适配),按实测多标: + # · 四条 DeepSeek 三线全通(chat + messages + responses) + # · kimi-k3 / minimax-m3 / minimax-m2.5 / qwen3.6~3.8 系列:chat + messages + # 反向证据(为什么必须逐条实测而不是"猜全支持"):把只认 chat 的 glm 系打到 + # /messages 上是 500 Internal server error,把 messages 线的模型打到 + # /responses 上是 401 "not supported for format openai" —— 都不像"选错格式", + # 所以网关改成按本表拦在入口(400 并说清该模型支持哪些线)。 models: - # ── OpenAI 兼容 /v1/chat/completions 线(全部经真实 key 实测 200)── + # ── OpenAI Chat Completions 线 ──────────────────────── - id: "glm-5.3-flash" display_name: "GLM-5.3-Flash" type: "text" + protocols: ["chat/completions"] - id: "glm-5.3" display_name: "GLM-5.3" type: "text" + protocols: ["chat/completions"] - id: "glm-5.2" display_name: "GLM-5.2" type: "text" + protocols: ["chat/completions"] - id: "glm-5.1" display_name: "GLM-5.1" type: "text" - - id: "kimi-k3" - display_name: "Kimi K3" - type: "text" + protocols: ["chat/completions"] - id: "kimi-k2.7-code" display_name: "Kimi K2.7 Code" type: "text" + protocols: ["chat/completions"] # 上游硬约束:该模型**只接受 temperature=1**,传 0 / 0.7 等一律 400 # "invalid temperature: only 1 is allowed for this model"。 # 写死后网关无论收到什么温度都按 1 发(前端温度滑块、API 传参、一键测速探针 @@ -95,44 +131,120 @@ providers: - id: "kimi-k2.6" display_name: "Kimi K2.6" type: "text" + protocols: ["chat/completions"] - id: "longcat-2.0" display_name: "LongCat 2.0" type: "text" - - id: "deepseek-v4-pro" - display_name: "DeepSeek V4 Pro" + protocols: ["chat/completions"] + - id: "mimo-v2.5" + display_name: "MiMo V2.5" type: "text" - - id: "deepseek-v4-flash" - display_name: "DeepSeek V4 Flash" + protocols: ["chat/completions"] + - id: "mimo-v2.5-pro" + display_name: "MiMo V2.5 Pro" type: "text" - - id: "deepseek-flash" - display_name: "DeepSeek Flash" + protocols: ["chat/completions"] + - id: "hy4-preview" + display_name: "Hy4 Preview" type: "text" + protocols: ["chat/completions"] + - id: "hy3" + display_name: "Hy3" + type: "text" + protocols: ["chat/completions"] + # ── Chat + Messages 两条线都实测可用 ───────────────── + - id: "kimi-k3" + display_name: "Kimi K3" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "minimax-m3" + display_name: "MiniMax M3" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "minimax-m2.5" + display_name: "MiniMax M2.5" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "qwen3.8-max" + display_name: "Qwen3.8 Max" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "qwen3.8-flash" + display_name: "Qwen3.8 Flash" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "qwen3.7-max" + display_name: "Qwen3.7 Max" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "qwen3.7-plus" + display_name: "Qwen3.7 Plus" + type: "text" + protocols: ["chat/completions", "messages"] + - id: "qwen3.6-plus" + display_name: "Qwen3.6 Plus" + type: "text" + protocols: ["chat/completions", "messages"] + # ── 三条线全通(实测 chat / messages / responses 均 200, + # 非文档线另经流式 + 工具调用复核) ──────────────── - id: "deepseek-v4.1-flash" display_name: "DeepSeek V4.1 Flash" type: "text" vision: true # 2026-09-18 实测确认多模态:data:image/png;base64 红/蓝/绿三色均识别正确; # 同一带图请求打到纯文本模型 deepseek-v4-flash 会被上游 400 拒(对照有效)。 + protocols: ["chat/completions", "messages", "responses"] + - id: "deepseek-v4-pro" + display_name: "DeepSeek V4 Pro" + type: "text" + protocols: ["chat/completions", "messages", "responses"] + - id: "deepseek-v4-flash" + display_name: "DeepSeek V4 Flash" + type: "text" + protocols: ["chat/completions", "messages", "responses"] - id: "deepseek-v4-flash-vision-exp" display_name: "DeepSeek V4 Flash Vision Exp" type: "text" - vision: true # 官方标 Vision,支持图像转 token 输入(带图请求可走路由链首) - - id: "mimo-v2.5" - display_name: "MiMo V2.5" + vision: true # 官方标 Vision,支持图像转 token 输入 + protocols: ["chat/completions", "messages", "responses"] + # ── Anthropic Messages 线(文档主推这条)───────────── + - id: "minimax-m2.7" + display_name: "MiniMax M2.7" type: "text" - - id: "mimo-v2.5-pro" - display_name: "MiMo V2.5 Pro" + # 实测:messages 线 200,chat / responses 都是 500 → 只标 messages + protocols: ["messages"] + - id: "union-alpha" + display_name: "Union Alpha Free" type: "text" - - id: "hy4-preview" - display_name: "Hy4 Preview" + # 文档标的 messages 线,但实测上游回 400「Model is unavailable.」 + # (限时免费模型,额度/可用性由上游控制)。先按文档收录并标注真实线, + # 调用时上游的不可用会原样回给客户端(网关不替它编文案)。 + protocols: ["messages"] + # ── OpenAI Responses 线 ─────────────────────────────── + - id: "grok-4.6" + display_name: "Grok 4.6" type: "text" - - id: "hy3" - display_name: "Hy3" + # 实测:responses 200;chat / messages 都是 401 "not supported for format" + protocols: ["responses"] + - id: "gpt-5.6-luna" + display_name: "GPT 5.6 Luna" type: "text" + protocols: ["responses"] + - id: "muse-spark-1.3-contributor" + display_name: "Muse Spark 1.3 Contributor" + type: "text" + # 实测三条线都 403 DataPolicyError:该模型会拿数据训练,需要先在 + # OpenCode 控制台**显式同意数据条款**才能调用(当前工作区未同意)。 + # 先收录并标好线:老板去控制台点了同意就能直接用,不用再改网关。 + protocols: ["responses"] + - id: "muse-spark-1.2-contributor" + display_name: "Muse Spark 1.2 Contributor" + type: "text" + protocols: ["responses"] server: host: "127.0.0.1" port: 8808 timeout: 120 - retry: 3 # 直接用模型 id(不走路由)时重试 3 次(含首次);路由别名用各自 route.retry(默认1) + retry: 3 # 同一模型的重试次数(含首次)。第二十七轮: 路由别名已删除,无 route.retry # ── 运维开关 ── health_auth: false # /health/* 是否也需鉴权(默认否,留给探活) hot_reload: true # 监听 gateway.yaml mtime 自动重载(改配置免重启) @@ -195,36 +307,10 @@ download: auth: session_ttl_minutes: 1440 -# ── 路由别名 + 按优先级 fallback ────────────────────── -# model 填别名(如 wjd),按数组顺序逐个尝试: -# 第 1 个失败/超时/重试耗尽 → 第 2 个 … 全失败才 502。 -# 每个路由可设 timeout(单模型超时秒)、retry(切下一个前该模型重试次数)。 -# 优先级从左到右(第一个最高)。老板决策:wjd 链路 = 全部文本模型(已含 OpenCode Go 17 个)。 -routes: - wjd: - timeout: 60 # 单模型 60s 没出结果就切下一个(避免 20×120s 挂死) - retry: 1 # 每个模型只试 1 次,失败立即切下一个 - # 注意(老板决策):wjd 链路 = 全部文本模型;新增/现有文本模型都必须加进这里。 - models: - # 供应商按「默认主力」在前,其余按供应商分组;运行时仍按可用性+速度实时重排。 - - "ark-code-latest" # 1) 火山方舟 - - "agnes-3.0-flash" # 2) Agnes 3.0 - - "agnes-2.5-flash" # 3) Agnes 2.5 - # ── 4) OpenCode Go(17 个文本模型,全进 wjd)── - - "glm-5.3-flash" # 4.1) GLM-5.3-Flash - - "glm-5.3" # 4.2) GLM-5.3 - - "glm-5.2" # 4.3) GLM-5.2 - - "glm-5.1" # 4.4) GLM-5.1 - - "kimi-k3" # 4.5) Kimi K3 - - "kimi-k2.7-code" # 4.6) Kimi K2.7 Code - - "kimi-k2.6" # 4.7) Kimi K2.6 - - "longcat-2.0" # 4.8) LongCat 2.0 - - "deepseek-v4-pro" # 4.9) DeepSeek V4 Pro - - "deepseek-v4-flash" # 4.10) DeepSeek V4 Flash - - "deepseek-flash" # 4.11) DeepSeek Flash - - "deepseek-v4.1-flash" # 4.12) DeepSeek V4.1 Flash - - "deepseek-v4-flash-vision-exp" # 4.13) DeepSeek V4 Flash Vision Exp - - "mimo-v2.5" # 4.14) MiMo V2.5 - - "mimo-v2.5-pro" # 4.15) MiMo V2.5 Pro - - "hy4-preview" # 4.16) Hy4 Preview - - "hy3" # 4.17) Hy3 +# ── 路由别名 / fallback 链:已整体删除 ────────────────── +# 第二十七轮(老板 2026-09-18 决策):**路由别名功能整体删除**,不再有 routes 段。 +# 原因:三线直传落地后"入站格式 = 出站格式",模型 id 必须与协议线一一对应; +# 再套一层"自动挑最快模型"的别名,会把调用方选定的格式/能力(视觉、工具、协议线) +# 悄悄换掉。现在调用方直接填真实模型 id;要换模型就换 id,网关不再替谁做主。 +# 老配置里残留的 routes 段会被忽略(不会让进程起不来),日志里有 warning 可查。 +# 一键测速(Models 页每个模型后面的)不受影响,仍然逐个模型实测延迟。 diff --git a/docs/frontend-design.md b/docs/frontend-design.md index aa4b608..fc05908 100644 --- a/docs/frontend-design.md +++ b/docs/frontend-design.md @@ -5,6 +5,10 @@ --- +> ⚠️ **第二十七轮(2026-09-18)**:路由别名 `wjd` / 多模型 fallback 已**整体删除**。 +> 本文里凡是提到「路由别名」「降级链」「fallback_used」的地方,都是当时的设计记录, +> 现状请以 `README.md` 的「三线直传」章节为准。 + ## 1. 背景与目标 网关是一个 OpenAI 兼容的多供应商 LLM 统一接入层(Agnes / 火山方舟),已具备完整 REST 能力(对话/图像/视频/批量/模型/指标/日志/健康)。当前**只能通过 curl / 二次开发调用,没有可视化界面**。 diff --git a/frontend/package-lock.json b/frontend/package-lock.json index ad4b326..1397f6e 100644 --- a/frontend/package-lock.json +++ b/frontend/package-lock.json @@ -8,7 +8,6 @@ "name": "api-gateway-web", "version": "0.1.0", "dependencies": { - "@element-plus/icons-vue": "^2.3.1", "dompurify": "^3.4.15", "echarts": "^6.1.0", "element-plus": "^2.9.1", @@ -23,10 +22,21 @@ "devDependencies": { "@vitejs/plugin-vue": "^5.2.1", "typescript": "^5.7.2", + "unplugin-auto-import": "^0.19.0", + "unplugin-vue-components": "^0.28.0", "vite": "^6.0.7", "vue-tsc": "^2.2.12" } }, + "node_modules/@antfu/utils": { + "version": "0.7.10", + "resolved": "http://mirrors.tencentyun.com/npm/@antfu/utils/-/utils-0.7.10.tgz", + "integrity": "sha512-+562v9k4aI80m1+VuMHehNJWLOFjBnXn3tdOitzD0il5b7smkSBal4+a3oKiQTbrwMmN/TBUMDvbdoWDehgOww==", + "dev": true, + "funding": { + "url": "https://github.com/sponsors/antfu" + } + }, "node_modules/@babel/helper-string-parser": { "version": "7.29.7", "resolved": "http://mirrors.tencentyun.com/npm/@babel/helper-string-parser/-/helper-string-parser-7.29.7.tgz", @@ -564,11 +574,50 @@ "url": "https://github.com/sponsors/kazupon" } }, + "node_modules/@jridgewell/gen-mapping": { + "version": "0.3.13", + "resolved": "http://mirrors.tencentyun.com/npm/@jridgewell/gen-mapping/-/gen-mapping-0.3.13.tgz", + "integrity": "sha512-2kkt/7niJ6MgEPxF0bYdQ6etZaA+fQvDcLKckhy1yIQOzaoKjBBjSj63/aLVjYE3qhRt5dvM+uUyfCg6UKCBbA==", + "dev": true, + "dependencies": { + "@jridgewell/sourcemap-codec": "^1.5.0", + "@jridgewell/trace-mapping": "^0.3.24" + } + }, + "node_modules/@jridgewell/remapping": { + "version": "2.3.5", + "resolved": "http://mirrors.tencentyun.com/npm/@jridgewell/remapping/-/remapping-2.3.5.tgz", + "integrity": "sha512-LI9u/+laYG4Ds1TDKSJW2YPrIlcVYOwi2fUC6xB43lueCjgxV4lffOCZCtYFiH6TNOX+tQKXx97T4IKHbhyHEQ==", + "dev": true, + "dependencies": { + "@jridgewell/gen-mapping": "^0.3.5", + "@jridgewell/trace-mapping": "^0.3.24" + } + }, + "node_modules/@jridgewell/resolve-uri": { + "version": "3.1.2", + "resolved": "http://mirrors.tencentyun.com/npm/@jridgewell/resolve-uri/-/resolve-uri-3.1.2.tgz", + "integrity": "sha512-bRISgCIjP20/tbWSPWMEi54QVPRZExkuD9lJL+UIxUKtwVJA8wW1Trb1jMs1RFXo1CBTNZ/5hpC9QvmKWdopKw==", + "dev": true, + "engines": { + "node": ">=6.0.0" + } + }, "node_modules/@jridgewell/sourcemap-codec": { "version": "1.6.0", "resolved": "http://mirrors.tencentyun.com/npm/@jridgewell/sourcemap-codec/-/sourcemap-codec-1.6.0.tgz", "integrity": "sha512-T7jf+5zgsZHwNJ4lvQ7/aezbyk0nNX+zJVWpmHA7VYsEx7a7qr5Rg5IbtJFqkgze5Y2sruq1RUY8Q837Od7iFw==" }, + "node_modules/@jridgewell/trace-mapping": { + "version": "0.3.31", + "resolved": "http://mirrors.tencentyun.com/npm/@jridgewell/trace-mapping/-/trace-mapping-0.3.31.tgz", + "integrity": "sha512-zzNR+SdQSDJzc8joaeP8QQoCQr8NuYx2dIIytl1QeBEZHJ9uW6hebsrYgbz8hJwUQao3TWCMtmfV8Nu1twOLAw==", + "dev": true, + "dependencies": { + "@jridgewell/resolve-uri": "^3.1.0", + "@jridgewell/sourcemap-codec": "^1.4.14" + } + }, "node_modules/@napi-rs/lzma-linux-x64-gnu": { "version": "1.5.1", "resolved": "http://mirrors.tencentyun.com/npm/@napi-rs/lzma-linux-x64-gnu/-/lzma-linux-x64-gnu-1.5.1.tgz", @@ -585,6 +634,41 @@ "node": "^22.20 || ^24.12 || >=25" } }, + "node_modules/@nodelib/fs.scandir": { + "version": "2.1.5", + "resolved": "http://mirrors.tencentyun.com/npm/@nodelib/fs.scandir/-/fs.scandir-2.1.5.tgz", + "integrity": "sha512-vq24Bq3ym5HEQm2NKCr3yXDwjc7vTsEThRDnkp2DK9p1uqLR+DHurm/NOTo0KG7HYHU7eppKZj3MyqYuMBf62g==", + "dev": true, + "dependencies": { + "@nodelib/fs.stat": "2.0.5", + "run-parallel": "^1.1.9" + }, + "engines": { + "node": ">= 8" + } + }, + "node_modules/@nodelib/fs.stat": { + "version": "2.0.5", + "resolved": "http://mirrors.tencentyun.com/npm/@nodelib/fs.stat/-/fs.stat-2.0.5.tgz", + "integrity": "sha512-RkhPPp2zrqDAQA/2jNhnztcPAlv64XdhIp7a7454A5ovI7Bukxgt7MX7udwAu3zg1DcpPU0rz3VV1SeaqvY4+A==", + "dev": true, + "engines": { + "node": ">= 8" + } + }, + "node_modules/@nodelib/fs.walk": { + "version": "1.2.8", + "resolved": "http://mirrors.tencentyun.com/npm/@nodelib/fs.walk/-/fs.walk-1.2.8.tgz", + "integrity": "sha512-oGB+UxlgWcgQkgwo8GcEGwemoTFt3FIO9ababBmaGwXIoBKZ+GTy0pP185beGg7Llih/NSHSV2XAs1lnznocSg==", + "dev": true, + "dependencies": { + "@nodelib/fs.scandir": "2.1.5", + "fastq": "^1.6.0" + }, + "engines": { + "node": ">= 8" + } + }, "node_modules/@popperjs/core": { "name": "@sxzz/popperjs-es", "version": "2.11.8", @@ -595,6 +679,28 @@ "url": "https://opencollective.com/popperjs" } }, + "node_modules/@rollup/pluginutils": { + "version": "5.4.0", + "resolved": "http://mirrors.tencentyun.com/npm/@rollup/pluginutils/-/pluginutils-5.4.0.tgz", + "integrity": "sha512-MfPp06CjRLfXQ3wY0R8vJDYBy/MvVcc9OulEfR0B8Iv9ko+GCNaRZ+EpJYFl27LhKsZK0o420sYCRHCjfCgeUg==", + "dev": true, + "dependencies": { + "@types/estree": "^1.0.0", + "estree-walker": "^2.0.2", + "picomatch": "^4.0.2" + }, + "engines": { + "node": ">=14.0.0" + }, + "peerDependencies": { + "rollup": "^1.20.0||^2.0.0||^3.0.0||^4.0.0" + }, + "peerDependenciesMeta": { + "rollup": { + "optional": true + } + } + }, "node_modules/@rollup/rollup-android-arm-eabi": { "version": "4.63.1", "resolved": "http://mirrors.tencentyun.com/npm/@rollup/rollup-android-arm-eabi/-/rollup-android-arm-eabi-4.63.1.tgz", @@ -1152,12 +1258,49 @@ "vue": "^3.5.0" } }, + "node_modules/acorn": { + "version": "8.18.0", + "resolved": "http://mirrors.tencentyun.com/npm/acorn/-/acorn-8.18.0.tgz", + "integrity": "sha512-lGq+9yr1/GuAWaVYIHRjvvySG5/4VfKIvC8EWxStPdcDh/Ka7FG3twP6v4d5BkravUilhIAsG4Qj83t02LWUPQ==", + "dev": true, + "bin": { + "acorn": "bin/acorn" + }, + "engines": { + "node": ">=0.4.0" + } + }, "node_modules/alien-signals": { "version": "1.0.13", "resolved": "http://mirrors.tencentyun.com/npm/alien-signals/-/alien-signals-1.0.13.tgz", "integrity": "sha512-OGj9yyTnJEttvzhTUWuscOvtqxq5vrhF7vL9oS0xJ2mK0ItPYP1/y+vCFebfxoEyAz0++1AIwJ5CMr+Fk3nDmg==", "dev": true }, + "node_modules/anymatch": { + "version": "3.1.3", + "resolved": "http://mirrors.tencentyun.com/npm/anymatch/-/anymatch-3.1.3.tgz", + "integrity": "sha512-KMReFUr0B4t+D+OBkjR3KYqvocp2XaSzO55UcB6mgQMd3KbcE+mWTyvVV7D/zsdEbNnV6acZUutkiHQXvTr1Rw==", + "dev": true, + "dependencies": { + "normalize-path": "^3.0.0", + "picomatch": "^2.0.4" + }, + "engines": { + "node": ">= 8" + } + }, + "node_modules/anymatch/node_modules/picomatch": { + "version": "2.3.2", + "resolved": "http://mirrors.tencentyun.com/npm/picomatch/-/picomatch-2.3.2.tgz", + "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==", + "dev": true, + "engines": { + "node": ">=8.6" + }, + "funding": { + "url": "https://github.com/sponsors/jonschlinkert" + } + }, "node_modules/async-validator": { "version": "4.2.5", "resolved": "http://mirrors.tencentyun.com/npm/async-validator/-/async-validator-4.2.5.tgz", @@ -1169,6 +1312,18 @@ "integrity": "sha512-3oSeUO0TMV67hN1AmbXsK4yaqU7tjiHlbxRDZOpH0KW9+CeX4bRAaX0Anxt0tx2MrpRpWwQaPwIlISEJhYU5Pw==", "dev": true }, + "node_modules/binary-extensions": { + "version": "2.3.0", + "resolved": "http://mirrors.tencentyun.com/npm/binary-extensions/-/binary-extensions-2.3.0.tgz", + "integrity": "sha512-Ceh+7ox5qe7LJuLHoY0feh3pHuUDHAcRUeyL2VYghZwfpkNIy/+8Ocg0a3UuSoYzavmylwuLWQOf3hl0jjMMIw==", + "dev": true, + "engines": { + "node": ">=8" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, "node_modules/brace-expansion": { "version": "2.1.4", "resolved": "http://mirrors.tencentyun.com/npm/brace-expansion/-/brace-expansion-2.1.4.tgz", @@ -1178,6 +1333,48 @@ "balanced-match": "^1.0.0" } }, + "node_modules/braces": { + "version": "3.0.3", + "resolved": "http://mirrors.tencentyun.com/npm/braces/-/braces-3.0.3.tgz", + "integrity": "sha512-yQbXgO/OSZVD2IsiLlro+7Hf6Q18EJrKSEsdoMzKePKXct3gvD8oLcOQdIzGupr5Fj+EDe8gO/lxc1BzfMpxvA==", + "dev": true, + "dependencies": { + "fill-range": "^7.1.1" + }, + "engines": { + "node": ">=8" + } + }, + "node_modules/chokidar": { + "version": "3.6.0", + "resolved": "http://mirrors.tencentyun.com/npm/chokidar/-/chokidar-3.6.0.tgz", + "integrity": "sha512-7VT13fmjotKpGipCW9JEQAusEPE+Ei8nl6/g4FBAmIm0GOOLMua9NDDo/DWp0ZAxCr3cPq5ZpBqmPAQgDda2Pw==", + "dev": true, + "dependencies": { + "anymatch": "~3.1.2", + "braces": "~3.0.2", + "glob-parent": "~5.1.2", + "is-binary-path": "~2.1.0", + "is-glob": "~4.0.1", + "normalize-path": "~3.0.0", + "readdirp": "~3.6.0" + }, + "engines": { + "node": ">= 8.10.0" + }, + "funding": { + "url": "https://paulmillr.com/funding/" + }, + "optionalDependencies": { + "fsevents": "~2.3.2" + } + }, + "node_modules/confbox": { + "version": "0.1.8", + "resolved": "http://mirrors.tencentyun.com/npm/confbox/-/confbox-0.1.8.tgz", + "integrity": "sha512-RMtmw0iFkeR4YV+fUOSucriAQNb9g8zFR52MWCtl+cCZOFRNL6zeB395vPzFhEjjn4fMxXudmELnl/KF/WrK6w==", + "dev": true + }, "node_modules/csstype": { "version": "3.2.3", "resolved": "http://mirrors.tencentyun.com/npm/csstype/-/csstype-3.2.3.tgz", @@ -1194,6 +1391,23 @@ "integrity": "sha512-e/1zu3xH5MQryN2zdVaF0OrdNLUbvWxzMbi+iNA6Bky7l1RoP8a2fIbRocyHclXt/arDrrR6lL3TqFD9pMQTsg==", "dev": true }, + "node_modules/debug": { + "version": "4.4.3", + "resolved": "http://mirrors.tencentyun.com/npm/debug/-/debug-4.4.3.tgz", + "integrity": "sha512-RGwwWnwQvkVfavKVt22FGLw+xYSdzARwm0ru6DhTVA3umU5hZc28V3kO4stgYryrTlLpuvgI9GiijltAjNbcqA==", + "dev": true, + "dependencies": { + "ms": "^2.1.3" + }, + "engines": { + "node": ">=6.0" + }, + "peerDependenciesMeta": { + "supports-color": { + "optional": true + } + } + }, "node_modules/dompurify": { "version": "3.4.15", "resolved": "http://mirrors.tencentyun.com/npm/dompurify/-/dompurify-3.4.15.tgz", @@ -1288,11 +1502,55 @@ "@esbuild/win32-x64": "0.25.12" } }, + "node_modules/escape-string-regexp": { + "version": "5.0.0", + "resolved": "http://mirrors.tencentyun.com/npm/escape-string-regexp/-/escape-string-regexp-5.0.0.tgz", + "integrity": "sha512-/veY75JbMK4j1yjvuUxuVsiS/hr/4iHs9FTT6cgTexxdE0Ly/glccBAkloH/DofkjRbZU3bnoj38mOmhkZ0lHw==", + "dev": true, + "license": "MIT", + "engines": { + "node": ">=12" + }, + "funding": { + "url": "https://github.com/sponsors/sindresorhus" + } + }, "node_modules/estree-walker": { "version": "2.0.2", "resolved": "http://mirrors.tencentyun.com/npm/estree-walker/-/estree-walker-2.0.2.tgz", "integrity": "sha512-Rfkk/Mp/DL7JVje3u18FxFujQlTNR2q6QfMSMB7AvCBx91NGj/ba3kCfza0f6dVDbw7YlRf/nDrn7pQrCCyQ/w==" }, + "node_modules/exsolve": { + "version": "1.1.1", + "resolved": "http://mirrors.tencentyun.com/npm/exsolve/-/exsolve-1.1.1.tgz", + "integrity": "sha512-9U/jZUgjnSGyntRr6y5Muu1MJcwFl6kPu7k8qLF0IMNfLqvw0NZ4nnVDq0RVoZ0RvCyumib4Ez3KYrVfilrw+g==", + "dev": true + }, + "node_modules/fast-glob": { + "version": "3.3.3", + "resolved": "http://mirrors.tencentyun.com/npm/fast-glob/-/fast-glob-3.3.3.tgz", + "integrity": "sha512-7MptL8U0cqcFdzIzwOTHoilX9x5BrNqye7Z/LuC7kCMRio1EMSyqRK3BEAUD7sXRq4iT4AzTVuZdhgQ2TCvYLg==", + "dev": true, + "dependencies": { + "@nodelib/fs.stat": "^2.0.2", + "@nodelib/fs.walk": "^1.2.3", + "glob-parent": "^5.1.2", + "merge2": "^1.3.0", + "micromatch": "^4.0.8" + }, + "engines": { + "node": ">=8.6.0" + } + }, + "node_modules/fastq": { + "version": "1.20.3", + "resolved": "http://mirrors.tencentyun.com/npm/fastq/-/fastq-1.20.3.tgz", + "integrity": "sha512-XKv5nnLs6nLF71NgiKJLIZFLkPyIEuOselLG7ujZnGrRfQK8HpvY+WqKhAJUAdLomwVHErVS4LfxFlPq0/FTAw==", + "dev": true, + "dependencies": { + "reusify": "^1.0.4" + } + }, "node_modules/fdir": { "version": "6.5.0", "resolved": "http://mirrors.tencentyun.com/npm/fdir/-/fdir-6.5.0.tgz", @@ -1310,6 +1568,18 @@ } } }, + "node_modules/fill-range": { + "version": "7.1.1", + "resolved": "http://mirrors.tencentyun.com/npm/fill-range/-/fill-range-7.1.1.tgz", + "integrity": "sha512-YsGpe3WHLK8ZYi4tWDg2Jy3ebRz2rXowDxnld4bkQB00cc/1Zw9AWnC0i9ztDJitivtQvaI9KaLyKrc+hBW0yg==", + "dev": true, + "dependencies": { + "to-regex-range": "^5.0.1" + }, + "engines": { + "node": ">=8" + } + }, "node_modules/fsevents": { "version": "2.3.3", "resolved": "http://mirrors.tencentyun.com/npm/fsevents/-/fsevents-2.3.3.tgz", @@ -1324,6 +1594,18 @@ "node": "^8.16.0 || ^10.6.0 || >=11.0.0" } }, + "node_modules/glob-parent": { + "version": "5.1.2", + "resolved": "http://mirrors.tencentyun.com/npm/glob-parent/-/glob-parent-5.1.2.tgz", + "integrity": "sha512-AOIgSQCepiJYwP3ARnGx+5VnTu2HBYdzbGP45eLw1vr3zB3vZLeyed1sC9hnbcOc9/SrMyM5RPQrkGz4aS9Zow==", + "dev": true, + "dependencies": { + "is-glob": "^4.0.1" + }, + "engines": { + "node": ">= 6" + } + }, "node_modules/he": { "version": "1.2.0", "resolved": "http://mirrors.tencentyun.com/npm/he/-/he-1.2.0.tgz", @@ -1341,6 +1623,70 @@ "node": ">=12.0.0" } }, + "node_modules/is-binary-path": { + "version": "2.1.0", + "resolved": "http://mirrors.tencentyun.com/npm/is-binary-path/-/is-binary-path-2.1.0.tgz", + "integrity": "sha512-ZMERYes6pDydyuGidse7OsHxtbI7WVeUEozgR/g7rd0xUimYNlvZRE/K2MgZTjWy725IfelLeVcEM97mmtRGXw==", + "dev": true, + "dependencies": { + "binary-extensions": "^2.0.0" + }, + "engines": { + "node": ">=8" + } + }, + "node_modules/is-extglob": { + "version": "2.1.1", + "resolved": "http://mirrors.tencentyun.com/npm/is-extglob/-/is-extglob-2.1.1.tgz", + "integrity": "sha512-SbKbANkN603Vi4jEZv49LeVJMn4yGwsbzZworEoyEiutsN3nJYdbO36zfhGJ6QEDpOZIFkDtnq5JRxmvl3jsoQ==", + "dev": true, + "engines": { + "node": ">=0.10.0" + } + }, + "node_modules/is-glob": { + "version": "4.0.3", + "resolved": "http://mirrors.tencentyun.com/npm/is-glob/-/is-glob-4.0.3.tgz", + "integrity": "sha512-xelSayHH36ZgE7ZWhli7pW34hNbNl8Ojv5KVmkJD4hBdD3th8Tfk9vYasLM+mXWOZhFkgZfxhLSnrwRr4elSSg==", + "dev": true, + "dependencies": { + "is-extglob": "^2.1.1" + }, + "engines": { + "node": ">=0.10.0" + } + }, + "node_modules/is-number": { + "version": "7.0.0", + "resolved": "http://mirrors.tencentyun.com/npm/is-number/-/is-number-7.0.0.tgz", + "integrity": "sha512-41Cifkg6e8TylSpdtTpeLVMqvSBEVzTttHvERD741+pnZ8ANv0004MRL43QKPDlK9cGvNp6NZWZUBlbGXYxxng==", + "dev": true, + "engines": { + "node": ">=0.12.0" + } + }, + "node_modules/js-tokens": { + "version": "9.0.1", + "resolved": "http://mirrors.tencentyun.com/npm/js-tokens/-/js-tokens-9.0.1.tgz", + "integrity": "sha512-mxa9E9ITFOt0ban3j6L5MpjwegGz6lBQmM1IJkWeBZGcMxto50+eWdjC/52xDbS2vy0k7vIMK0Fe2wfL9OQSpQ==", + "dev": true + }, + "node_modules/local-pkg": { + "version": "0.5.1", + "resolved": "http://mirrors.tencentyun.com/npm/local-pkg/-/local-pkg-0.5.1.tgz", + "integrity": "sha512-9rrA30MRRP3gBD3HTGnC6cDFpaE1kVDWxWgqWJUN0RvDNAo+Nz/9GxB+nHOH0ifbVFy0hSA1V6vFDvnx54lTEQ==", + "dev": true, + "dependencies": { + "mlly": "^1.7.3", + "pkg-types": "^1.2.1" + }, + "engines": { + "node": ">=14" + }, + "funding": { + "url": "https://github.com/sponsors/antfu" + } + }, "node_modules/lodash": { "version": "4.18.1", "resolved": "http://mirrors.tencentyun.com/npm/lodash/-/lodash-4.18.1.tgz", @@ -1394,6 +1740,40 @@ "resolved": "http://mirrors.tencentyun.com/npm/memoize-one/-/memoize-one-6.0.0.tgz", "integrity": "sha512-rkpe71W0N0c0Xz6QD0eJETuWAJGnJ9afsl1srmwPrI+yBCkge5EycXXbYRyvL29zZVUWQCY7InPRCv3GDXuZNw==" }, + "node_modules/merge2": { + "version": "1.4.1", + "resolved": "http://mirrors.tencentyun.com/npm/merge2/-/merge2-1.4.1.tgz", + "integrity": "sha512-8q7VEgMJW4J8tcfVPy8g09NcQwZdbwFEqhe/WZkoIzjn/3TGDwtOCYtXGxA3O8tPzpczCCDgv+P2P5y00ZJOOg==", + "dev": true, + "engines": { + "node": ">= 8" + } + }, + "node_modules/micromatch": { + "version": "4.0.8", + "resolved": "http://mirrors.tencentyun.com/npm/micromatch/-/micromatch-4.0.8.tgz", + "integrity": "sha512-PXwfBhYu0hBCPw8Dn0E+WDYb7af3dSLVWKi3HGv84IdF4TyFoC0ysxFd0Goxw7nSv4T/PzEJQxsYsEiFCKo2BA==", + "dev": true, + "dependencies": { + "braces": "^3.0.3", + "picomatch": "^2.3.1" + }, + "engines": { + "node": ">=8.6" + } + }, + "node_modules/micromatch/node_modules/picomatch": { + "version": "2.3.2", + "resolved": "http://mirrors.tencentyun.com/npm/picomatch/-/picomatch-2.3.2.tgz", + "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==", + "dev": true, + "engines": { + "node": ">=8.6" + }, + "funding": { + "url": "https://github.com/sponsors/jonschlinkert" + } + }, "node_modules/minimatch": { "version": "9.0.9", "resolved": "http://mirrors.tencentyun.com/npm/minimatch/-/minimatch-9.0.9.tgz", @@ -1409,6 +1789,24 @@ "url": "https://github.com/sponsors/isaacs" } }, + "node_modules/mlly": { + "version": "1.8.2", + "resolved": "http://mirrors.tencentyun.com/npm/mlly/-/mlly-1.8.2.tgz", + "integrity": "sha512-d+ObxMQFmbt10sretNDytwt85VrbkhhUA/JBGm1MPaWJ65Cl4wOgLaB1NYvJSZ0Ef03MMEU/0xpPMXUIQ29UfA==", + "dev": true, + "dependencies": { + "acorn": "^8.16.0", + "pathe": "^2.0.3", + "pkg-types": "^1.3.1", + "ufo": "^1.6.3" + } + }, + "node_modules/ms": { + "version": "2.1.3", + "resolved": "http://mirrors.tencentyun.com/npm/ms/-/ms-2.1.3.tgz", + "integrity": "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA==", + "dev": true + }, "node_modules/muggle-string": { "version": "0.4.1", "resolved": "http://mirrors.tencentyun.com/npm/muggle-string/-/muggle-string-0.4.1.tgz", @@ -1432,6 +1830,15 @@ "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1" } }, + "node_modules/normalize-path": { + "version": "3.0.0", + "resolved": "http://mirrors.tencentyun.com/npm/normalize-path/-/normalize-path-3.0.0.tgz", + "integrity": "sha512-6eZs5Ls3WtCisHWp9S2GUy8dqkpGi4BVSz3GaqiE6ezub0512ESztXUwUB6C6IKbQkY2Pnb/mD4WYojCRwcwLA==", + "dev": true, + "engines": { + "node": ">=0.10.0" + } + }, "node_modules/normalize-wheel-es": { "version": "1.2.0", "resolved": "http://mirrors.tencentyun.com/npm/normalize-wheel-es/-/normalize-wheel-es-1.2.0.tgz", @@ -1443,6 +1850,12 @@ "integrity": "sha512-b7uo2UCUOYZcnF/3ID0lulOJi/bafxa1xPe7ZPsammBSpjSWQkjNxlt635YGS2MiR9GjvuXCtz2emr3jbsz98g==", "dev": true }, + "node_modules/pathe": { + "version": "2.0.3", + "resolved": "http://mirrors.tencentyun.com/npm/pathe/-/pathe-2.0.3.tgz", + "integrity": "sha512-WUjGcAqP1gQacoQe+OBJsFA7Ld4DyXuUIjZ5cc75cLHvJ7dtNsTugphxIADwspS+AraAUePCKrSVtPLFj/F88w==", + "dev": true + }, "node_modules/picocolors": { "version": "1.1.1", "resolved": "http://mirrors.tencentyun.com/npm/picocolors/-/picocolors-1.1.1.tgz", @@ -1481,6 +1894,17 @@ } } }, + "node_modules/pkg-types": { + "version": "1.3.1", + "resolved": "http://mirrors.tencentyun.com/npm/pkg-types/-/pkg-types-1.3.1.tgz", + "integrity": "sha512-/Jm5M4RvtBFVkKWRu2BLUTNP8/M2a+UwuAX+ae4770q1qVGtfjG+WTCupoZixokjmHiry8uI+dlY8KXYV5HVVQ==", + "dev": true, + "dependencies": { + "confbox": "^0.1.8", + "mlly": "^1.7.4", + "pathe": "^2.0.1" + } + }, "node_modules/postcss": { "version": "8.5.28", "resolved": "http://mirrors.tencentyun.com/npm/postcss/-/postcss-8.5.28.tgz", @@ -1508,6 +1932,76 @@ "node": "^10 || ^12 || >=14" } }, + "node_modules/quansync": { + "version": "0.2.11", + "resolved": "http://mirrors.tencentyun.com/npm/quansync/-/quansync-0.2.11.tgz", + "integrity": "sha512-AifT7QEbW9Nri4tAwR5M/uzpBuqfZf+zwaEM/QkzEjj7NBuFD2rBuy0K3dE+8wltbezDV7JMA0WfnCPYRSYbXA==", + "dev": true, + "funding": [ + { + "type": "individual", + "url": "https://github.com/sponsors/antfu" + }, + { + "type": "individual", + "url": "https://github.com/sponsors/sxzz" + } + ] + }, + "node_modules/queue-microtask": { + "version": "1.2.3", + "resolved": "http://mirrors.tencentyun.com/npm/queue-microtask/-/queue-microtask-1.2.3.tgz", + "integrity": "sha512-NuaNSa6flKT5JaSYQzJok04JzTL1CA6aGhv5rfLW3PgqA+M2ChpZQnAC8h8i4ZFkBS8X5RqkDBHA7r4hej3K9A==", + "dev": true, + "funding": [ + { + "type": "github", + "url": "https://github.com/sponsors/feross" + }, + { + "type": "patreon", + "url": "https://www.patreon.com/feross" + }, + { + "type": "consulting", + "url": "https://feross.org/support" + } + ] + }, + "node_modules/readdirp": { + "version": "3.6.0", + "resolved": "http://mirrors.tencentyun.com/npm/readdirp/-/readdirp-3.6.0.tgz", + "integrity": "sha512-hOS089on8RduqdbhvQ5Z37A0ESjsqz6qnRcffsMU3495FuTdqSm+7bhJ29JvIOsBDEEnan5DPu9t3To9VRlMzA==", + "dev": true, + "dependencies": { + "picomatch": "^2.2.1" + }, + "engines": { + "node": ">=8.10.0" + } + }, + "node_modules/readdirp/node_modules/picomatch": { + "version": "2.3.2", + "resolved": "http://mirrors.tencentyun.com/npm/picomatch/-/picomatch-2.3.2.tgz", + "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==", + "dev": true, + "engines": { + "node": ">=8.6" + }, + "funding": { + "url": "https://github.com/sponsors/jonschlinkert" + } + }, + "node_modules/reusify": { + "version": "1.1.0", + "resolved": "http://mirrors.tencentyun.com/npm/reusify/-/reusify-1.1.0.tgz", + "integrity": "sha512-g6QUff04oZpHs0eG5p83rFLhHeV00ug/Yf9nZM6fLeUrPguBTkTQOdpAWWspMh55TZfVQDPaN3NQJfbVRAxdIw==", + "dev": true, + "engines": { + "iojs": ">=1.0.0", + "node": ">=0.10.0" + } + }, "node_modules/rollup": { "version": "4.63.1", "resolved": "http://mirrors.tencentyun.com/npm/rollup/-/rollup-4.63.1.tgz", @@ -1553,6 +2047,35 @@ "fsevents": "~2.3.2" } }, + "node_modules/run-parallel": { + "version": "1.2.0", + "resolved": "http://mirrors.tencentyun.com/npm/run-parallel/-/run-parallel-1.2.0.tgz", + "integrity": "sha512-5l4VyZR86LZ/lDxZTR6jqL8AFE2S0IFLMP26AbjsLVADxHdhB/c0GUsH+y39UfCi3dzz8OlQuPmnaJOMoDHQBA==", + "dev": true, + "funding": [ + { + "type": "github", + "url": "https://github.com/sponsors/feross" + }, + { + "type": "patreon", + "url": "https://www.patreon.com/feross" + }, + { + "type": "consulting", + "url": "https://feross.org/support" + } + ], + "dependencies": { + "queue-microtask": "^1.2.2" + } + }, + "node_modules/scule": { + "version": "1.3.0", + "resolved": "http://mirrors.tencentyun.com/npm/scule/-/scule-1.3.0.tgz", + "integrity": "sha512-6FtHJEvt+pVMIB9IBY+IcCJ6Z5f1iQnytgyfKMhDKgmzYG+TeH/wx1y3l27rshSbLiSanrR9ffZDrEsmjlQF2g==", + "dev": true + }, "node_modules/source-map-js": { "version": "1.2.1", "resolved": "http://mirrors.tencentyun.com/npm/source-map-js/-/source-map-js-1.2.1.tgz", @@ -1561,6 +2084,18 @@ "node": ">=0.10.0" } }, + "node_modules/strip-literal": { + "version": "2.1.1", + "resolved": "http://mirrors.tencentyun.com/npm/strip-literal/-/strip-literal-2.1.1.tgz", + "integrity": "sha512-631UJ6O00eNGfMiWG78ck80dfBab8X6IVFB51jZK5Icd7XAs60Z5y7QdSd/wGIklnWvRbUNloVzhOKKmutxQ6Q==", + "dev": true, + "dependencies": { + "js-tokens": "^9.0.1" + }, + "funding": { + "url": "https://github.com/sponsors/antfu" + } + }, "node_modules/tinyglobby": { "version": "0.2.17", "resolved": "http://mirrors.tencentyun.com/npm/tinyglobby/-/tinyglobby-0.2.17.tgz", @@ -1577,6 +2112,18 @@ "url": "https://github.com/sponsors/SuperchupuDev" } }, + "node_modules/to-regex-range": { + "version": "5.0.1", + "resolved": "http://mirrors.tencentyun.com/npm/to-regex-range/-/to-regex-range-5.0.1.tgz", + "integrity": "sha512-65P7iz6X5yEr1cwcgvQxbbIw7Uk3gOy5dIdtZ4rDveLqhrdJP+Li/Hx6tyK0NEb+2GCyneCMJiGqrADCSNk8sQ==", + "dev": true, + "dependencies": { + "is-number": "^7.0.0" + }, + "engines": { + "node": ">=8.0" + } + }, "node_modules/tslib": { "version": "2.3.0", "resolved": "http://mirrors.tencentyun.com/npm/tslib/-/tslib-2.3.0.tgz", @@ -1596,6 +2143,175 @@ "node": ">=14.17" } }, + "node_modules/ufo": { + "version": "1.6.4", + "resolved": "http://mirrors.tencentyun.com/npm/ufo/-/ufo-1.6.4.tgz", + "integrity": "sha512-JFNbkD1Svwe0KvGi8GOeLcP4kAWQ609twvCdcHxq1oSL8svv39ZuSvajcD8B+5D0eL4+s1Is2D/O6KN3qcTeRA==", + "dev": true + }, + "node_modules/unimport": { + "version": "3.14.6", + "resolved": "http://mirrors.tencentyun.com/npm/unimport/-/unimport-3.14.6.tgz", + "integrity": "sha512-CYvbDaTT04Rh8bmD8jz3WPmHYZRG/NnvYVzwD6V1YAlvvKROlAeNDUBhkBGzNav2RKaeuXvlWYaa1V4Lfi/O0g==", + "dev": true, + "dependencies": { + "@rollup/pluginutils": "^5.1.4", + "acorn": "^8.14.0", + "escape-string-regexp": "^5.0.0", + "estree-walker": "^3.0.3", + "fast-glob": "^3.3.3", + "local-pkg": "^1.0.0", + "magic-string": "^0.30.17", + "mlly": "^1.7.4", + "pathe": "^2.0.1", + "picomatch": "^4.0.2", + "pkg-types": "^1.3.0", + "scule": "^1.3.0", + "strip-literal": "^2.1.1", + "unplugin": "^1.16.1" + } + }, + "node_modules/unimport/node_modules/confbox": { + "version": "0.3.1", + "resolved": "http://mirrors.tencentyun.com/npm/confbox/-/confbox-0.3.1.tgz", + "integrity": "sha512-cKUSoKa8YxFZZSmraVi7onONx3amu77ngK3kGpsYHDH7drPwCRkQE1RYMPlLRrMtnciRj274XNRxcHxnKmDSnA==", + "dev": true + }, + "node_modules/unimport/node_modules/estree-walker": { + "version": "3.0.3", + "resolved": "http://mirrors.tencentyun.com/npm/estree-walker/-/estree-walker-3.0.3.tgz", + "integrity": "sha512-7RUKfXgSMMkzt6ZuXmqapOurLGPPfgj6l9uRZ7lRGolvk0y2yocc35LdcxKC5PQZdn2DMqioAQ2NoWcrTKmm6g==", + "dev": true, + "dependencies": { + "@types/estree": "^1.0.0" + } + }, + "node_modules/unimport/node_modules/local-pkg": { + "version": "1.2.1", + "resolved": "http://mirrors.tencentyun.com/npm/local-pkg/-/local-pkg-1.2.1.tgz", + "integrity": "sha512-++gUqRDEvcnN6Zhqrr+y/CkVEHhlrR96vZn3nZZPYzMcBUyBtTKzB9NadClFIsIVSsu+3i9tfk/erqy9kAmt7Q==", + "dev": true, + "dependencies": { + "mlly": "^1.7.4", + "pkg-types": "^2.3.0", + "quansync": "^0.2.11" + }, + "engines": { + "node": ">=14" + }, + "funding": { + "url": "https://github.com/sponsors/antfu" + } + }, + "node_modules/unimport/node_modules/local-pkg/node_modules/pkg-types": { + "version": "2.3.3", + "resolved": "http://mirrors.tencentyun.com/npm/pkg-types/-/pkg-types-2.3.3.tgz", + "integrity": "sha512-j/lCFdcppV0JxWpCEITdbDltBxPP6cHT+yNJ6Go2OgoSA9518X847X9z0p6LtA4Nc16+eQzCZjRrWanTGvHJ5w==", + "dev": true, + "dependencies": { + "confbox": "^0.3.1", + "exsolve": "^1.1.1", + "pathe": "^2.0.3" + } + }, + "node_modules/unimport/node_modules/unplugin": { + "version": "1.16.1", + "resolved": "http://mirrors.tencentyun.com/npm/unplugin/-/unplugin-1.16.1.tgz", + "integrity": "sha512-4/u/j4FrCKdi17jaxuJA0jClGxB1AvU2hw/IuayPc4ay1XGaJs/rbb4v5WKwAjNifjmXK9PIFyuPiaK8azyR9w==", + "dev": true, + "dependencies": { + "acorn": "^8.14.0", + "webpack-virtual-modules": "^0.6.2" + }, + "engines": { + "node": ">=14.0.0" + } + }, + "node_modules/unplugin": { + "version": "2.3.11", + "resolved": "http://mirrors.tencentyun.com/npm/unplugin/-/unplugin-2.3.11.tgz", + "integrity": "sha512-5uKD0nqiYVzlmCRs01Fhs2BdkEgBS3SAVP6ndrBsuK42iC2+JHyxM05Rm9G8+5mkmRtzMZGY8Ct5+mliZxU/Ww==", + "dev": true, + "dependencies": { + "@jridgewell/remapping": "^2.3.5", + "acorn": "^8.15.0", + "picomatch": "^4.0.3", + "webpack-virtual-modules": "^0.6.2" + }, + "engines": { + "node": ">=18.12.0" + } + }, + "node_modules/unplugin-auto-import": { + "version": "0.19.0", + "resolved": "http://mirrors.tencentyun.com/npm/unplugin-auto-import/-/unplugin-auto-import-0.19.0.tgz", + "integrity": "sha512-W97gTDEWu/L1EcKCXY5Ni8bsMW1E9kv12wYQv3mYpd7zcFctXYlLKsqeva6sbCQbzS8t9AG/XdU5/WkEJKPlFw==", + "dev": true, + "dependencies": { + "@antfu/utils": "^0.7.10", + "@rollup/pluginutils": "^5.1.3", + "local-pkg": "^0.5.1", + "magic-string": "^0.30.15", + "picomatch": "^4.0.2", + "unimport": "^3.14.5", + "unplugin": "^2.1.0" + }, + "engines": { + "node": ">=14" + }, + "funding": { + "url": "https://github.com/sponsors/antfu" + }, + "peerDependencies": { + "@nuxt/kit": "^3.2.2", + "@vueuse/core": "*" + }, + "peerDependenciesMeta": { + "@nuxt/kit": { + "optional": true + }, + "@vueuse/core": { + "optional": true + } + } + }, + "node_modules/unplugin-vue-components": { + "version": "0.28.0", + "resolved": "http://mirrors.tencentyun.com/npm/unplugin-vue-components/-/unplugin-vue-components-0.28.0.tgz", + "integrity": "sha512-jiTGtJ3JsRFBjgvyilfrX7yUoGKScFgbdNw+6p6kEXU+Spf/rhxzgvdfuMcvhCcLmflB/dY3pGQshYBVGOUx7Q==", + "dev": true, + "dependencies": { + "@antfu/utils": "^0.7.10", + "@rollup/pluginutils": "^5.1.4", + "chokidar": "^3.6.0", + "debug": "^4.4.0", + "fast-glob": "^3.3.2", + "local-pkg": "^0.5.1", + "magic-string": "^0.30.15", + "minimatch": "^9.0.5", + "mlly": "^1.7.3", + "unplugin": "^2.1.0" + }, + "engines": { + "node": ">=14" + }, + "funding": { + "url": "https://github.com/sponsors/antfu" + }, + "peerDependencies": { + "@babel/parser": "^7.15.8", + "@nuxt/kit": "^3.2.2", + "vue": "2 || 3" + }, + "peerDependenciesMeta": { + "@babel/parser": { + "optional": true + }, + "@nuxt/kit": { + "optional": true + } + } + }, "node_modules/vite": { "version": "6.4.3", "resolved": "http://mirrors.tencentyun.com/npm/vite/-/vite-6.4.3.tgz", @@ -1775,6 +2491,12 @@ "typescript": ">=5.0.0" } }, + "node_modules/webpack-virtual-modules": { + "version": "0.6.2", + "resolved": "http://mirrors.tencentyun.com/npm/webpack-virtual-modules/-/webpack-virtual-modules-0.6.2.tgz", + "integrity": "sha512-66/V2i5hQanC51vBQKPH4aI8NMAcBW59FVBs+rC7eGHupMyfn34q7rZIE+ETlJ+XTevqfUhVVBgSUNSW2flEUQ==", + "dev": true + }, "node_modules/zrender": { "version": "6.1.0", "resolved": "http://mirrors.tencentyun.com/npm/zrender/-/zrender-6.1.0.tgz", diff --git a/frontend/package.json b/frontend/package.json index 5cf45e9..6799265 100644 --- a/frontend/package.json +++ b/frontend/package.json @@ -12,7 +12,6 @@ "check:i18n": "node scripts/check-i18n.mjs" }, "dependencies": { - "@element-plus/icons-vue": "^2.3.1", "dompurify": "^3.4.15", "echarts": "^6.1.0", "element-plus": "^2.9.1", @@ -27,6 +26,8 @@ "devDependencies": { "@vitejs/plugin-vue": "^5.2.1", "typescript": "^5.7.2", + "unplugin-auto-import": "^0.19.0", + "unplugin-vue-components": "^0.28.0", "vite": "^6.0.7", "vue-tsc": "^2.2.12" } diff --git a/frontend/src/lib/chat.ts b/frontend/src/lib/chat.ts index e6158a5..801b921 100644 --- a/frontend/src/lib/chat.ts +++ b/frontend/src/lib/chat.ts @@ -7,7 +7,6 @@ export interface Meta { served: string provider: string requested: string - fallback: boolean } export interface Usage { @@ -25,7 +24,7 @@ export interface ChatMessage { imgThumbs?: string[] // 第十三轮:缩略图 data URI —— 落服务端历史时用它替代原图 // (服务端单条 512KB,原图 base64 会被截断成坏图) reasoning?: string // 推理内容(reasoning_content),单独折叠展示 - meta?: Meta // 网关元信息(served/provider/fallback) + meta?: Meta // 网关元信息(served/provider) usage?: Usage cost?: number // estimated_cost_cny(仅非流式) streaming?: boolean // 正在流式输出中 diff --git a/frontend/src/locales/en.ts b/frontend/src/locales/en.ts index 83ed69c..54ea049 100644 --- a/frontend/src/locales/en.ts +++ b/frontend/src/locales/en.ts @@ -19,10 +19,10 @@ export default { control: 'Console', more: 'More', subdashboard: 'Gateway overview · refreshed every 8s', - subchat: 'Vision-ready · route wjd picks the fastest provider', + subchat: 'Vision-ready · multi-provider direct routing', subimages: 'Params on the left, history on the right · downloads proxied by the gateway', subvideos: '720P · 4–12s · async polling', - submodels: 'Providers / models / route chains', + submodels: 'Providers / models / protocol lines', submetrics: 'Usage & cost · isolated per account', subkeys: 'Up to 3 per account · stored in plaintext, copyable anytime', subusers: 'Accounts & roles · admin can inspect others’ logs and assets', @@ -86,8 +86,6 @@ export default { dashboard: { modelCount: 'Models', providers: 'Providers', - routes: 'Route aliases', - fallback: 'Fallbacks', inflight: 'In-flight', callsMix: 'Calls ok/fail', provShare: 'Provider success share', @@ -97,7 +95,6 @@ export default { circuit: 'Circuit', modelCountShort: 'Models', keySet: 'Key set', - routeAlias: 'Route aliases', name: 'Name', modelChain: 'Model chain', gatewayInfo: 'Gateway info', @@ -114,7 +111,6 @@ export default { requests: 'Total requests', successRate: 'Success rate', avgLatency: 'Avg latency', - fallbackShort: 'Fallbacks', noteOkFail: 'ok {ok} · fail {fail}', noteFail: '{n} failed', noteP95: 'highest P95 {v}s', @@ -127,22 +123,21 @@ export default { noModels: 'No calls recorded yet', share: 'Share', live: 'Live', - routeChain: 'Availability + speed first', + byProtocol: 'Protocol lines', + byProtocolSub: 'Models per line', + modelCountN: '{n} models', }, metrics: { uptime: 'Uptime', inflight: 'In-flight/peak', - fallback: 'Fallbacks', promptTokens: 'Prompt tokens', completionTokens: 'Completion tokens', callsTotal: 'Total calls (ok/fail)', modelCalls: 'Calls per model (success / fail)', provShare: 'Provider success share', - routeVsProvider: 'Route vs provider (success / fail / fallback)', latency: 'Model latency P50 / P95 (s)', success: 'Success', fail: 'Fail', - fallback2: 'Fallback', ok: 'ok', none: 'None', circuitStatus: 'Circuit breaker status', @@ -165,6 +160,7 @@ export default { failRateNote: 'failure rate {v}%', snapshot: 'Live snapshot', circuitCard: 'Circuit breakers', + byProvider: 'Calls by provider (success / fail)', }, logs: { all: 'All', @@ -226,20 +222,17 @@ export default { systemShort: 'System', visionReady: 'Vision ready · images supported', visionOff: 'This model does not accept images', - chainLabel: '{id} ({chain})', emptyChain: 'empty', phDefault: 'Default', phOptional: 'Optional', totalTokens: 'Total {n}', uploadedImage: 'Uploaded image', reasoningTag: 'Reasoning', - routeGroup: 'Routes (auto fallback)', textModels: 'Text models', params: 'Parameters', clear: 'Clear', generating: 'Generating…', - empty: 'Start a conversation with the gateway (any text model or route wjd)', - fallback: 'Fallback · by {provider}', + empty: 'Start a conversation with the gateway (pick a model below)', servedBy: 'by {provider}', requestedTo: 'requested {req} → served {served}', reasoning: 'Reasoning', @@ -266,6 +259,7 @@ export default { streaming: 'Streaming output', system: 'System prompt (prepended to each conversation)', loadModelsErr: 'Failed to load models: {msg}', + pickModel: 'Pick a model', }, images: { historyTitle: 'History', @@ -363,10 +357,8 @@ export default { loadErr: 'Failed to load videos', }, models: { - routeAliasTitle: 'Route aliases (auto fallback chain)', modelListTitle: 'Model list', name: 'Name', - chainLabel: 'Model chain (tried in order, auto fallback on failure)', providerCount: 'Providers', id: 'Model ID', type: 'Type', @@ -375,12 +367,10 @@ export default { text: 'Text', image: 'Image', video: 'Video', - route: 'Route', model: 'Models', textModels: 'Text models', imageModels: 'Image models', videoModels: 'Video models', - routes: 'Route aliases', providers: 'Providers', loadErr: 'Failed to load models', capability: 'Capability', @@ -391,10 +381,8 @@ export default { providerCard: 'Providers', keyOk: 'Key configured', keyMissing: 'Key missing', - routeCount: '{n} routes', chainCount: '{n} models', providerCount2: '{n} providers', - noRoute: 'No route alias configured', speedtest: 'Test speed', testing: 'Probing latency of every text model…', testingShort: 'Testing…', @@ -403,6 +391,16 @@ export default { failed: 'Failed', speedtestDone: 'Speed test done, {n} models', speedtestFail: 'Speed test failed', + speedtestHint: 'Latency probed per protocol line', + textModelCount: '{n} text models', + // Round 27: the supported-formats column (three protocol lines). Line names are the + // vendors' own endpoint names, so they are not translated. + protocols: 'Supported formats', + protocolHint: 'Which lines this model accepts: chat/completions (OpenAI-compatible), messages (Anthropic), responses (OpenAI Responses). Any other format is rejected with 400 at the gateway edge — the gateway never translates between formats and never silently swaps in another model', + protocolNa: 'N/A', + protocolNaHint: 'Image / video models use the dedicated /v1/images/generations and /v1/videos endpoints and are not tied to the three protocol lines', + protocolHintRow: 'Supported formats: {lines} ', + probedLine: 'probed via {p}', }, history: { saveFail: 'History could not be saved to the gateway (it may be lost on refresh) — check your session or network', @@ -435,7 +433,7 @@ export default { logoutBtn: 'Sign out', chatDefaults: 'Chat default parameters', defsTip: 'Used as defaults for a new chat (already-adjusted params are untouched).', - defaultModel: 'Default model / route', + defaultModel: 'Default model', temperature: 'Temperature', topP: 'top_p', maxTokens: 'max_tokens (empty = gateway default)', @@ -448,7 +446,7 @@ export default { modelsCount: 'Models', userManagement: 'Account management (admin)', gwConfigBtn: 'Open', - gwConfigSub: 'Tune gateway / providers / routes / pricing — applied on save', + gwConfigSub: 'Tune gateway / providers / pricing — applied on save', gwConfig: 'Gateway config (admin)', userCol: 'User', roleCol: 'Role', @@ -504,38 +502,50 @@ export default { auth: 'Auth', authVal: 'Header: Authorization: Bearer ', modelPick: 'Model selection', - modelPickVal: 'model: wjd = route alias (auto-fallback by availability + speed, recommended); or a concrete model id to bypass fallback', + modelPickVal: 'model takes a real model id (e.g. agnes-3.0-flash). Each of the three lines forwards its own native protocol and the gateway never translates formats, so a wrong line returns 400 (listing the lines that model supports)', + formats: 'Supported formats', + formatsVal: 'Which lines a model accepts is listed in the Supported formats column on the Providers & Models page: chat/completions, messages, responses. Each line is forwarded verbatim (no field rewriting, no cross-format translation), so write your URL and body exactly as the vendor SDK expects', ctype: 'Request format', ctypeVal: 'application/json; chat supports SSE streaming (stream: true)', endpoints: 'Endpoints', colMethod: 'Method', colPath: 'Path', colDesc: 'Description', - epModels: 'Models and routes (with a vision flag)', + epModels: 'Models (with vision flag and supported protocol lines)', epChat: 'Chat completions (OpenAI compatible, streaming + image input)', + epMessages: 'Chat completions (Anthropic Messages format, forwarded verbatim)', + epResponses: 'Chat completions (OpenAI Responses format, forwarded verbatim)', epImages: 'Image generation / image-to-image (n must be 1; call repeatedly for more)', epVideoCreate: 'Video generation (async, returns video_id immediately)', epVideoPoll: 'Poll a video task (every ~5s; returns throttled when upstream rate-limits)', + epDownload: 'Download generated images / videos through the gateway (bypasses upstream hotlink protection and expiry; allow-listed hosts only)', epHistory: 'Read your own history (images / videos / chats)', epLogs: 'Your own call log', epMetrics: 'Metrics snapshot (incl. circuit breaker states)', epHealth: 'Health check (version / provider status)', exChat: 'Example · chat', + exMessages: 'Example · chat (Anthropic Messages format)', + exResponses: 'Example · chat (OpenAI Responses format)', exImg: 'Example · image generation', exVideo: 'Example · video generation + polling', exPy: 'Example · Python (OpenAI SDK compatible)', // Comments / sample payloads inside the code samples (mirror of zh.ts) sampleChat: 'Hello, introduce yourself', + sampleModelMessages: 'minimax-m2.7', + sampleModelResponses: 'grok-4.6', sampleImg: 'Neon city street at night', sampleVideo: 'Slow camera push-in', sampleStep1: 'submit a task (async, returns video_id immediately)', sampleStep2: 'poll every 5s (a throttled reply means wait, then retry)', notes: 'Notes', - noteVision: 'For image input use a vision model (agnes-3.0-flash, marked "Vision" in the list); with the wjd alias the gateway puts vision models first.', + noteVision: 'For image input use a vision model (agnes-3.0-flash, marked Vision in the list). The gateway no longer swaps models for you: a non-vision model receives the request as-is', + noteCross: 'No cross-format translation: which lines a model accepts is decided by the Supported formats column on the Providers & Models page (e.g. minimax-m2.7 only speaks messages, grok-4.6 only speaks responses). A wrong format is rejected with 400 at the gateway edge, listing the lines that model does support — your request is never silently rewritten, and no other model is swapped in', + noteConvId: 'Optional conversation_id (chat-style requests): reusing the same string for one conversation raises the upstream prompt-cache hit rate. It is a gateway-internal field and is never forwarded as-is — the gateway derives the real session id per account, so accounts never share a cache namespace; omit it and each call gets a random one', noteN1: 'Image n must be 1 (upstream limit) — call once per image; n≠1 is rejected with a clean 400.', noteVideo: 'Video seconds: 4–12 only; size is fixed at 720P; 6 aspect ratios (21:9/16:9/4:3/1:1/3:4/9:16); mode=text / keyframe / reference.', notePoll: 'Do not poll faster than 5s (upstream answers 429 "too frequent"; the gateway returns throttled so you can back off).', noteMedia: 'Assets (reference images / first-last frames / audio / chat images) accept data:image/...;base64,... or a public URL; data URIs avoid exposing the file, but keep each request body under 2MB.', + noteDownload: 'Upstream image / video URLs may expire or block hotlinking. To save a file, fetch /v1/media/download?url=... with your key and let the gateway retrieve it; URLs outside the allow-list are rejected with 400.', noteKey: 'If a key leaks, delete it here (401 immediately) or revoke all keys of the account from the Users page.', noteAudit: 'Every call is logged per account on the Logs page (you see yours; admins can inspect each account from the Users page).', copyBase: 'Copy URL', @@ -656,17 +666,13 @@ export default { addItem: 'Add', addModel: 'Add model', addProvider: 'Add provider', - addRoute: 'Add route alias', addPrice: 'Add pricing', delete: 'Delete', delProvider: 'Delete provider', - delRoute: 'Delete route', delConfirm: 'Delete "{name}"? It takes effect once you save.', needKey: 'Provider key cannot be empty', - needRoute: 'Route alias cannot be empty', dupKey: '"{name}" is duplicated — please rename', newProvider: '(new provider)', - newRoute: '(new alias)', unlimited: 'empty = unlimited', useServerTimeout: 'empty = server.timeout', mediaEmpty: 'empty = unsupported', @@ -683,19 +689,17 @@ export default { tabServerSub: 'Binding, timeout/retry, rate limits, CORS, session TTL and other runtime switches', tabProviders: 'Providers & models', tabProvidersSub: 'Key source, base URL, image/video endpoints and the model list per provider', - tabRoutes: 'Routes & circuit', - tabRoutesSub: 'Route aliases with their fallback chains (order = priority) plus the circuit breaker', tabLog: 'Logs & pricing', tabLogSub: 'Log level/format/capacity plus model pricing (CNY per million tokens)', tabDownload: 'Download proxy', tabDownloadSub: 'CDN hosts allowed for the download proxy and the per-file size limit', circuit: 'Provider circuit breaker', - h_circuit: 'A provider that fails to the threshold within the window is skipped for the cooldown; fallback takes over', + h_circuit: 'A provider failing to the threshold is cooled down; its requests get 503', f_c_enabled: 'Enable circuit breaker', f_c_failure_threshold: 'Failure threshold', f_c_cooldown: 'Cooldown (s)', f_c_window: 'Window (s)', - h_c_enabled: 'Off = never skip a failing provider (fallback still works)', + h_c_enabled: 'Off = never skip a failing provider', h_c_failure_threshold: 'Opens the circuit once failures reach this count in the window', h_c_cooldown: 'While open the provider is skipped; it half-opens automatically after the cooldown', h_c_window: 'Counts failures inside the window (not necessarily consecutive)', @@ -704,9 +708,9 @@ export default { f_port: 'Port', h_port: 'Production uses 8808 and Nginx points at it — change the proxy too', f_timeout: 'Upstream timeout (s)', - h_timeout: 'Per upstream request; a route-level timeout overrides it', + h_timeout: 'Per upstream request timeout', f_retry: 'Retries for direct model id', - h_retry: 'Retries (incl. first try) when calling a real model id instead of a route alias', + h_retry: 'Retries for the same model (incl. the first try)', f_health_auth: 'Require auth on /health', // vue-i18n parses {...} as a placeholder: literal braces must be written {'{'} / {'}'}, // otherwise the whole message fails to compile (and the UI using it renders nothing). @@ -719,7 +723,7 @@ export default { f_metrics_enabled: 'Record metrics', h_metrics_enabled: 'Off = /metrics and dashboard counters stop updating (calls still work)', f_rate_limit_rps: 'Global QPS per account', - h_rate_limit_rps: 'Empty = unlimited (provider/route level limits are configured separately)', + h_rate_limit_rps: 'Empty = unlimited (provider-level limits live in the provider entry)', f_max_body_size: 'Max request body (bytes)', h_max_body_size: '~2MB = 2000000; keep it generous if you send images', f_cors_origins: 'CORS allowed origins', @@ -755,17 +759,6 @@ export default { pModels: 'Models', h_pModels: 'Model ids must match upstream; vision = accepts image input (preferred for image chats)', mVision: 'vision', - routeHint: 'A route alias is a virtual model name (e.g. wjd): models are tried top-down and only a full failure returns 502. Order = priority.', - rName: 'Alias', - h_rName: 'Pass this as "model" when calling', - rTimeout: 'Per-model timeout (s)', - h_rTimeout: 'How long a single model may take before switching to the next one', - rRetry: 'Per-model retries', - h_rRetry: 'Retries before moving on (incl. first try); 1 = switch immediately', - rRate: 'Route QPS limit', - h_rRate: 'Empty = unlimited', - rModels: 'Fallback chain', - h_rModels: 'Tried top-down; available ids come from the providers above', f_o_log_level: 'Log level', h_o_log_level: 'DEBUG is noisy; INFO is fine for production', f_o_log_json: 'Structured JSON logs', diff --git a/frontend/src/locales/zh.ts b/frontend/src/locales/zh.ts index f80d932..1c2ab26 100644 --- a/frontend/src/locales/zh.ts +++ b/frontend/src/locales/zh.ts @@ -19,10 +19,10 @@ export default { control: '控制台', more: '更多', subdashboard: '网关运行概况 · 数据每 8 秒刷新', - subchat: '支持附图视觉 · 路由 wjd 自动选路', + subchat: '支持附图视觉 · 多供应商直连', subimages: '参数在左 · 历史在右 · 下载走网关代拉', subvideos: '720P · 时长 4–12 秒 · 异步轮询', - submodels: '供应商 / 模型 / 路由链', + submodels: '供应商 / 模型 / 协议线', submetrics: '用量与成本 · 按账号隔离', subkeys: '每个账号最多 3 把 · 明文保存,可反复复制', subusers: '账号与权限 · 管理员可查他人日志与资产', @@ -86,8 +86,6 @@ export default { dashboard: { modelCount: '模型数', providers: '供应商', - routes: '路由别名', - fallback: '成功降级', inflight: '当前并发', callsMix: '调用 ok/fail', provShare: '供应商成功调用份额', @@ -97,7 +95,6 @@ export default { circuit: '熔断', modelCountShort: '模型数', keySet: '已配 Key', - routeAlias: '路由别名', name: '名称', modelChain: '模型链', gatewayInfo: '网关信息', @@ -114,7 +111,6 @@ export default { requests: '请求总数', successRate: '成功率', avgLatency: '平均延迟', - fallbackShort: '成功降级', noteOkFail: 'ok {ok} · fail {fail}', noteFail: '失败 {n} 次', noteP95: 'P95 最高 {v}s', @@ -127,22 +123,21 @@ export default { noModels: '暂无调用记录', share: '份额', live: '实时', - routeChain: '可用性 + 速度优先链', + byProtocol: '协议线覆盖', + byProtocolSub: '每条线有多少模型', + modelCountN: '{n} 个模型', }, metrics: { uptime: '已运行', inflight: '在途/峰值', - fallback: '成功降级', promptTokens: 'Prompt tokens', completionTokens: 'Completion tokens', callsTotal: '总调用(ok/fail)', modelCalls: '各模型调用(成功 / 失败)', provShare: '供应商成功调用份额', - routeVsProvider: '路由 vs 供应商调用(成功 / 失败 / 降级)', latency: '模型延迟 P50 / P95(秒)', success: '成功', fail: '失败', - fallback2: '降级', ok: 'ok', none: '暂无', circuitStatus: '熔断器状态', @@ -165,6 +160,7 @@ export default { failRateNote: '失败率 {v}%', snapshot: '实时快照', circuitCard: '熔断器', + byProvider: '按供应商调用(成功 / 失败)', }, logs: { all: '全部', @@ -226,20 +222,17 @@ export default { systemShort: '系统提示', visionReady: '视觉能力已就绪 · 可附图', visionOff: '当前模型不支持图像输入', - chainLabel: '{id}({chain})', emptyChain: '空', phDefault: '默认', phOptional: '可选', totalTokens: '共 {n}', uploadedImage: '已上传图片', reasoningTag: '推理', - routeGroup: '路由(自动降级)', textModels: '文本模型', params: '参数', clear: '清空', generating: '正在生成…', - empty: '开始与网关对话(可选任意文本模型或路由 wjd)', - fallback: '降级 · 由 {provider} 响应', + empty: '开始与网关对话(在下方选一个模型)', servedBy: '由 {provider} 响应', requestedTo: '请求 {req} → 实际 {served}', reasoning: '推理过程(reasoning)', @@ -266,6 +259,7 @@ export default { streaming: '流式输出', system: '系统提示词(每条会话开头注入)', loadModelsErr: '加载模型列表失败:{msg}', + pickModel: '选择模型', }, images: { historyTitle: '生成历史', @@ -363,10 +357,8 @@ export default { loadErr: '加载视频失败', }, models: { - routeAliasTitle: '路由别名(自动降级链)', modelListTitle: '模型清单', name: '名称', - chainLabel: '模型链(依次尝试,失败自动降级)', providerCount: '供应商数', id: '模型 ID', type: '类型', @@ -375,12 +367,10 @@ export default { text: '文本', image: '图像', video: '视频', - route: '路由', model: '模型', textModels: '文本模型', imageModels: '图像模型', videoModels: '视频模型', - routes: '路由别名', providers: '供应商', loadErr: '加载模型失败', capability: '能力', @@ -391,10 +381,8 @@ export default { providerCard: '供应商', keyOk: 'Key 已配置', keyMissing: 'Key 未配置', - routeCount: '{n} 条路由', chainCount: '{n} 个模型', providerCount2: '{n} 家供应商', - noRoute: '没有配置路由别名', speedtest: '一键测速', testing: '正在测试全部文本模型延迟…', testingShort: '测速中…', @@ -403,6 +391,15 @@ export default { failed: '失败', speedtestDone: '测速完成,共 {n} 个模型', speedtestFail: '测速失败', + speedtestHint: '按模型自己的协议线探测延迟', + textModelCount: '{n} 个文本模型', + // 第二十七轮:支持的格式列(三条协议线)。格式名直接用各家标准端点名,不译。 + protocols: '支持的格式', + protocolHint: '该模型可以走哪几条线:chat/completions(OpenAI 兼容)、messages(Anthropic)、responses(OpenAI Responses)。不在列表里的格式会在网关入口直接 400 —— 网关不做跨格式转换,也绝不偷偷换个模型顶上', + protocolNa: '不适用', + protocolNaHint: '图像 / 视频模型走 /v1/images/generations、/v1/videos 专用端点,不涉及这三条协议线', + protocolHintRow: '支持的格式:{lines} ', + probedLine: '本次测速走 {p}', }, history: { saveFail: '历史未能保存到服务端(刷新后可能丢失),请检查登录状态或网络', @@ -435,7 +432,7 @@ export default { logoutBtn: '退出登录', chatDefaults: '对话默认参数', defsTip: '作为对话页初始值(本会话已改动的参数不受影响)。', - defaultModel: '默认模型 / 路由', + defaultModel: '默认模型', temperature: '温度 temperature', topP: 'top_p', maxTokens: 'max_tokens(留空用网关默认)', @@ -448,7 +445,7 @@ export default { modelsCount: '模型数', userManagement: '账号管理(管理员)', gwConfigBtn: '去配置', - gwConfigSub: '改网关参数 / 供应商 / 路由 / 定价 —— 保存即生效', + gwConfigSub: '改网关参数 / 供应商 / 定价 —— 保存即生效', gwConfig: '网关配置(管理员)', userCol: '账号', roleCol: '角色', @@ -504,38 +501,50 @@ export default { auth: '鉴权', authVal: '请求头 Authorization: Bearer <你的 API Key>', modelPick: '模型选择', - modelPickVal: 'model 填 wjd = 路由别名(按「可用 + 速度」自动降级,推荐);也可填具体模型 id 直连,不被降级影响', + modelPickVal: 'model 填真实模型 id(如 agnes-3.0-flash);三条线各按自家协议直传,网关不做跨格式转换 —— 选错线会 400(响应里会列出该模型支持哪些线)', + formats: '支持的格式', + formatsVal: '每个模型能用哪几条线,见「供应商与模型」页的「支持的格式」列:chat/completions、messages、responses。三条线各自原样转发(不改字段、不做跨格式翻译),因此接口地址与请求体按各家官方 SDK 写就行', ctype: '请求格式', ctypeVal: 'application/json;对话支持 SSE 流式(stream: true)', endpoints: '常用端点', colMethod: '方法', colPath: '路径', colDesc: '说明', - epModels: '模型与路由列表(含视觉能力标记)', + epModels: '模型列表(含视觉能力标记与支持的协议线)', epChat: '对话补全(OpenAI 兼容,支持流式与图片输入)', + epMessages: '对话补全(Anthropic Messages 格式,原样直传)', + epResponses: '对话补全(OpenAI Responses 格式,原样直传)', epImages: '图像生成 / 图生图(n 必须为 1,多张请多次调用)', epVideoCreate: '视频生成(异步,立即返回 video_id)', epVideoPoll: '查询视频任务(建议每 5s 一次;被上游限流时返回 throttled)', + epDownload: '代下载生成出来的图片 / 视频(绕过上游地址的防盗链与过期,白名单域名)', epHistory: '读取本人历史(图像 / 视频 / 对话)', epLogs: '本人调用日志', epMetrics: '指标快照(含熔断器状态)', epHealth: '健康检查(版本 / 供应商状态)', exChat: '示例 · 对话', + exMessages: '示例 · 对话(Anthropic Messages 格式)', + exResponses: '示例 · 对话(OpenAI Responses 格式)', exImg: '示例 · 图像生成', exVideo: '示例 · 视频生成 + 轮询', exPy: '示例 · Python(OpenAI SDK 兼容)', // 示例代码里的注释与示例文案(整段照抄时也该是当前语言) sampleChat: '你好,介绍一下你自己', + sampleModelMessages: 'minimax-m2.7', + sampleModelResponses: 'grok-4.6', sampleImg: '霓虹夜景的城市街道', sampleVideo: '镜头缓慢推进', sampleStep1: '提交任务(异步,立即返回 video_id)', sampleStep2: '每 5s 轮询一次(被限流会回 throttled,等一下再问)', notes: '注意事项', - noteVision: '带图对话请用支持视觉的模型(agnes-3.0-flash,列表里标「视觉」);填路由别名 wjd 时网关会自动把视觉模型排到链首。', + noteVision: '带图对话请用支持视觉的模型(agnes-3.0-flash,列表里标「视觉」);网关不再自动换模型 —— 选了不支持视觉的模型就照原样发上去,由上游决定怎么处理', + noteCross: '跨格式不转换:某个模型支持哪几条线,由「供应商与模型」页的「支持的格式」列决定(例:minimax-m2.7 只认 messages,grok-4.6 只认 responses)。用错格式网关会在入口直接 400 并列出它支持的线,不会偷偷翻译你的请求、也不会换个模型顶上', + noteConvId: '可选字段 conversation_id(对话类请求):同一段对话复用同一个串,上游的提示词缓存更容易命中(省钱省时)。它是网关内部字段,不会原样发给上游 —— 网关按账号派生真实会话 id,账号之间互不串缓存;不传则每次随机', noteN1: '图像 n 只能为 1(上游限制),要几张就调几次;n≠1 会被网关直接 400 拦下。', noteVideo: '视频 seconds 只支持 4–12,size 固定 720P,比例 6 种(21:9/16:9/4:3/1:1/3:4/9:16);mode=text 文生视频 / keyframe 首尾帧 / reference 参考图·音频。', notePoll: '轮询别快于 5s(上游会回 429「查询过于频繁」,网关会以 throttled 提示退避)。', noteMedia: '素材(参考图 / 首尾帧 / 参考音频 / 对话附图)可直接传 data:image/...;base64,... 或公网 URL;data URI 不必暴露素材,但单次请求体需 < 2MB。', + noteDownload: '上游返回的图片 / 视频地址可能有时效或防盗链。要落盘请走 /v1/media/download?url=...(需带上你的 Key),它由网关代为取回;地址不在白名单内会被 400 拒绝。', noteKey: 'Key 泄漏时到本页删除(立即 401),或用「用户管理」页一键吊销该账号全部 Key。', noteAudit: '所有调用按账号记入「日志」页(仅本人可见;admin 可在用户管理页看各账号)。', copyBase: '复制地址', @@ -656,17 +665,13 @@ export default { addItem: '添加', addModel: '添加模型', addProvider: '添加供应商', - addRoute: '添加路由别名', addPrice: '添加定价', delete: '删除', delProvider: '删除供应商', - delRoute: '删除路由', delConfirm: '删除「{name}」?保存后才真正写盘生效。', needKey: '供应商的 key 不能为空', - needRoute: '路由别名的名字不能为空', dupKey: '「{name}」重复了,请改一个名字', newProvider: '(新供应商)', - newRoute: '(新路由别名)', unlimited: '留空 = 不限', useServerTimeout: '留空 = 用 server.timeout', mediaEmpty: '留空 = 不支持', @@ -683,19 +688,17 @@ export default { tabServerSub: '监听、超时/重试、限流、CORS、会话时长等运行开关', tabProviders: '供应商与模型', tabProvidersSub: '供应商密钥来源、接口地址、图像/视频端点、模型清单', - tabRoutes: '路由与熔断', - tabRoutesSub: '路由别名与 fallback 模型链(顺序 = 优先级)+ 供应商熔断', tabLog: '日志与定价', tabLogSub: '日志级别/格式/容量 + 模型定价(人民币 / 百万 token)', tabDownload: '下载代理', tabDownloadSub: '允许网关代下载的 CDN 域名与单文件上限', circuit: '供应商熔断', - h_circuit: '某供应商窗口内失败到阈值就熔断冷却,期间跳过它让 fallback 顶上', + h_circuit: '某供应商窗口内失败到阈值就熔断冷却,期间它的请求直接 503', f_c_enabled: '启用熔断', f_c_failure_threshold: '触发阈值(次)', f_c_cooldown: '熔断冷却(秒)', f_c_window: '统计窗口(秒)', - h_c_enabled: '关掉 = 失败也不跳过(fallback 照常工作)', + h_c_enabled: '关掉 = 失败也不跳过', h_c_failure_threshold: '窗口内失败到该次数就熔断', h_c_cooldown: '熔断期间直接跳过该供应商,冷却结束后自动半开试探', h_c_window: '按"窗口内失败次数"计数(不是连续失败)', @@ -705,9 +708,9 @@ export default { f_port: '监听端口', h_port: '生产用 8808,且 Nginx 反代指向它 —— 改了要同步改反代', f_timeout: '上游超时(秒)', - h_timeout: '单次上游请求超时;路由自身写了 timeout 的以路由为准', + h_timeout: '单次上游请求超时', f_retry: '直连模型重试次数', - h_retry: '直接用模型 id(不走路由别名)时的重试次数,含首次', + h_retry: '同一模型的重试次数(含首次)', f_health_auth: '/health 需要鉴权', // 注意:vue-i18n 把 {...} 当占位符,花括号要么写成 {'{'} / {'}'},要么整条编译不过 —— // 编译不过的表现是"用到这条文案的整块界面不渲染",不报在文案上(踩过一次,见 check:i18n)。 @@ -720,7 +723,7 @@ export default { f_metrics_enabled: '记录调用指标', h_metrics_enabled: '关掉后 /metrics 与看板计数停更(不影响调用)', f_rate_limit_rps: '全局单账号 QPS 上限', - h_rate_limit_rps: '留空 = 不限(供应商级、路由级限流见各自配置)', + h_rate_limit_rps: '留空 = 不限(供应商级限流见供应商配置)', f_max_body_size: '请求体上限(字节)', h_max_body_size: '约 2MB = 2000000;带图对话别调太小', f_cors_origins: 'CORS 允许来源', @@ -757,18 +760,6 @@ export default { pModels: '模型清单', h_pModels: '模型 id 必须与上游一致;vision = 支持图像输入(带图对话会优先选它)', mVision: '视觉', - // ── 路由 ── - routeHint: '路由别名是给调用方用的"虚拟模型名"(如 wjd):按下面顺序逐个尝试,全失败才 502。数组顺序 = 优先级。', - rName: '别名', - h_rName: '调用时 model 填这个名字', - rTimeout: '单模型超时(秒)', - h_rTimeout: '某个模型多久没出结果就切下一个,避免整链挂死', - rRetry: '单模型重试次数', - h_rRetry: '切下一个之前的重试次数(含首次);想让链路快速切换就填 1', - rRate: '本路由 QPS 上限', - h_rRate: '留空 = 不限', - rModels: 'fallback 模型链', - h_rModels: '从上到下依次尝试;可用模型来自上面的供应商清单', // ── 可观测性 ── f_o_log_level: '日志级别', h_o_log_level: 'DEBUG 最啰嗦,生产 INFO 够用', diff --git a/frontend/src/main.ts b/frontend/src/main.ts index 278fef3..9d8dd85 100644 --- a/frontend/src/main.ts +++ b/frontend/src/main.ts @@ -1,10 +1,16 @@ import { createApp } from 'vue' import { createPinia } from 'pinia' -import ElementPlus from 'element-plus' -import 'element-plus/dist/index.css' +// 第二十八轮:element-plus 改按需引入(vite.config.ts 的 Components/AutoImport 负责模板里的 +// ),这里不再 `import ElementPlus` + `app.use(ElementPlus)`、不再引全量 index.css。 +// 仍然手动引的只有三样: +// ① 深色变量基座(A 皮肤用)—— 组件级样式依赖这三个 CSS 变量,必须全局在; +// ② ElMessage 的样式(Dialogs 里命令式调用,模板里没 ,resolver 扫不到); +// ③ ElMessageBox 的样式(同上,confirm() 弹窗)。 +// ElLoading / ElNotification 本项目未用,无需引(未来要用了记得补)。 import 'element-plus/theme-chalk/dark/css-vars.css' // 深色变量基座(A 皮肤用) +import 'element-plus/es/components/message/style/css' +import 'element-plus/es/components/message-box/style/css' import 'highlight.js/styles/github.css' -import * as ElementIcons from '@element-plus/icons-vue' import App from './App.vue' import router from './router' import { initTheme } from './lib/theme' @@ -15,9 +21,5 @@ initTheme() // mount 前应用主题(data-theme + dark class),避免闪色 const app = createApp(App) app.use(createPinia()) app.use(router) -app.use(ElementPlus) app.use(i18n) // 多语言 -for (const [name, comp] of Object.entries(ElementIcons)) { - app.component(name as string, comp as any) -} app.mount('#app') diff --git a/frontend/src/views/ApiKeys.vue b/frontend/src/views/ApiKeys.vue index 5bffd61..1d0f6b2 100644 --- a/frontend/src/views/ApiKeys.vue +++ b/frontend/src/views/ApiKeys.vue @@ -77,6 +77,9 @@
{{ t('apidoc.auth') }} {{ t('apidoc.authVal') }}
{{ t('apidoc.modelPick') }} {{ t('apidoc.modelPickVal') }}
+ +
{{ t('apidoc.formats') }} {{ t('apidoc.formatsVal') }}
{{ t('apidoc.ctype') }} {{ t('apidoc.ctypeVal') }}
@@ -141,12 +144,19 @@ const limit = ref(null) // null = 不限(admin) const base = window.location.origin // 端点表:与 app/main.py 的真实路由一一对应(别写成 README 里才有的东西) +// 第二十七轮:三条协议线并排 —— chat/completions(OpenAI 兼容,走专用路径)、 +// messages(Anthropic)、responses(OpenAI Responses)。后两条是「直传」:原始 JSON 进、 +// 上游响应原样出,网关不翻译字段;某个模型能不能用这条线,看「供应商与模型」页的 +// 「支持的格式」列(用错线会在入口 400 并列出该模型支持哪些线)。 const endpoints = computed(() => [ { method: 'GET', path: '/v1/models', desc: t('apidoc.epModels') }, { method: 'POST', path: '/v1/chat/completions', desc: t('apidoc.epChat') }, + { method: 'POST', path: '/v1/messages', desc: t('apidoc.epMessages') }, + { method: 'POST', path: '/v1/responses', desc: t('apidoc.epResponses') }, { method: 'POST', path: '/v1/images/generations', desc: t('apidoc.epImages') }, { method: 'POST', path: '/v1/videos', desc: t('apidoc.epVideoCreate') }, { method: 'GET', path: '/v1/videos/{video_id}?model=...', desc: t('apidoc.epVideoPoll') }, + { method: 'GET', path: '/v1/media/download?url=...', desc: t('apidoc.epDownload') }, { method: 'GET', path: '/history/{kind}', desc: t('apidoc.epHistory') }, { method: 'GET', path: '/logs', desc: t('apidoc.epLogs') }, { method: 'GET', path: '/metrics', desc: t('apidoc.epMetrics') }, @@ -154,8 +164,9 @@ const endpoints = computed(() => [ ]) const notes = computed(() => [ - t('apidoc.noteVision'), t('apidoc.noteN1'), t('apidoc.noteVideo'), - t('apidoc.notePoll'), t('apidoc.noteMedia'), t('apidoc.noteKey'), t('apidoc.noteAudit'), + t('apidoc.noteVision'), t('apidoc.noteCross'), t('apidoc.noteConvId'), t('apidoc.noteN1'), + t('apidoc.noteVideo'), t('apidoc.notePoll'), t('apidoc.noteMedia'), t('apidoc.noteDownload'), + t('apidoc.noteKey'), t('apidoc.noteAudit'), ]) // 示例里的 Key 用占位符 $GW_KEY(不把真 Key 渲染进文档;本页的 Key 明文只在上方列表里回显) @@ -166,7 +177,15 @@ const examples = computed(() => { { title: t('apidoc.exChat'), code: `curl ${base}/v1/chat/completions \\ ${H} \\ - -d '{"model":"wjd","messages":[{"role":"user","content":"${t('apidoc.sampleChat')}"}]}'` }, + -d '{"model":"agnes-3.0-flash","messages":[{"role":"user","content":"${t('apidoc.sampleChat')}"}]}'` }, + { title: t('apidoc.exMessages'), code: +`curl ${base}/v1/messages \\ + ${H} \\ + -d '{"model":"${t('apidoc.sampleModelMessages')}","max_tokens":64,"messages":[{"role":"user","content":"${t('apidoc.sampleChat')}"}]}'` }, + { title: t('apidoc.exResponses'), code: +`curl ${base}/v1/responses \\ + ${H} \\ + -d '{"model":"${t('apidoc.sampleModelResponses')}","input":"${t('apidoc.sampleChat')}","max_output_tokens":64}'` }, { title: t('apidoc.exImg'), code: `curl ${base}/v1/images/generations \\ ${H} \\ @@ -185,7 +204,7 @@ curl "${base}/v1/videos/task_xxx?model=agnes-video-2.5-flash" \\ `from openai import OpenAI c = OpenAI(base_url="${base}/v1", api_key="wjd-xxxxxxxx") -r = c.chat.completions.create(model="wjd", messages=[{"role": "user", "content": "${t('apidoc.sampleChat')}"}]) +r = c.chat.completions.create(model="agnes-3.0-flash", messages=[{"role": "user", "content": "${t('apidoc.sampleChat')}"}]) print(r.choices[0].message.content)` }, ] }) diff --git a/frontend/src/views/Chat.vue b/frontend/src/views/Chat.vue index 24e4cd4..7d3e81b 100644 --- a/frontend/src/views/Chat.vue +++ b/frontend/src/views/Chat.vue @@ -104,15 +104,12 @@
- +
- - + + {{ t('chat.servedBy', { provider: msg.meta.provider }) }} - - - - + {{ msg.meta.served }} · {{ msg.cost != null ? `¥${msg.cost}` : '' }} @@ -199,15 +196,7 @@ - - - - - {{ chainLabel(r) }} - - - - + @@ -289,36 +278,36 @@ const router = useRouter() const { t } = useI18n() // ── 模型选项 ── -const routes = ref([]) +// 第二十七轮: 路由别名(wjd)功能已整体删除 —— /v1/models 里只剩真实模型, +// 调用方选哪个模型就是哪个(网关不再"自动挑最快模型")。 const textModels = ref([]) async function loadModels() { try { const j = await api('/v1/models') const data: any[] = j.data || [] - routes.value = data.filter((x) => x.type === 'route') - textModels.value = data.filter((x) => x.type !== 'route') + textModels.value = data.filter((x) => x.type !== 'image' && x.type !== 'video') + // 默认模型:优先 agnes-3.0-flash(旗舰),否则用列表第一个。 + // 同时兜住"本地存着旧模型 id(比如已删除的 wjd)"的情况 —— 认不出来就换默认, + // 免得下拉框空着、发出去却被网关 400。 + if (!textModels.value.some((m) => m.id === params.model)) { + const def = textModels.value.find((m) => m.id === 'agnes-3.0-flash') + params.model = def ? def.id : (textModels.value[0]?.id || '') + } } catch (e: any) { if (e && e.status === 401) { router.push('/login'); return } ElMessage.error(t('chat.loadModelsErr', { msg: (e && e.message) || e })) } } -function chainLabel(r: any) { - const chain = (r.models || []).join(' → ') - // 括号样式(全角/半角)也交给 i18n:英文里用半角更自然 - return t('chat.chainLabel', { id: r.id, chain: chain || t('chat.emptyChain') }) -} // ── 视觉能力(第十三轮)── // 网关 /v1/models 会回 vision 字段(Agnes 3.0 Flash = true,其余文本模型默认 false)。 -// 带图请求走路由别名时网关会把视觉模型排到链首;选了**不支持视觉的真实模型**时 -// 这里提前告警并提供一键切换,避免"发了图但模型根本没看"。 +// 第二十七轮: 路由链删除后网关不再"把视觉模型排到链首" —— 选了不支持视觉的模型时, +// 这里提前告警并提供一键切换,避免"发了图但模型根本没看"。 const visionModels = computed(() => textModels.value.filter((m) => m.vision).map((m) => m.id)) const visionOk = computed(() => { - const m = routes.value.find((r) => r.id === params.model) - if (m) return !!m.vision || !!(m.models || []).some((id: string) => visionModels.value.includes(id)) - const t2 = textModels.value.find((x) => x.id === params.model) - return !!t2?.vision + const m = textModels.value.find((x) => x.id === params.model) + return !!m?.vision }) const firstVisionModel = computed(() => visionModels.value[0] || '') function useVisionModel() { @@ -330,7 +319,8 @@ function modelLabel(m: any) { // ── 参数与状态 ── const params = reactive({ - model: 'wjd', + // 空串 = 还没拿到模型列表;loadModels() 里填入默认模型(第二十七轮: 不再默认 wjd) + model: '', temperature: 0.7, top_p: 1, max_tokens: null as number | null, @@ -654,12 +644,11 @@ async function runOnce(input: string) { } reply.usage = j?.usage || undefined reply.cost = j?.estimated_cost_cny ?? undefined - if (j?.route || j?.provider) { + if (j?.provider || j?.served_model) { reply.meta = { served: j.served_model || j.model || params.model, provider: j.provider_name || j.provider, requested: j.model || params.model, - fallback: !!(j.route && j.route.fallback_used), } } } catch (e: any) { @@ -687,7 +676,6 @@ async function runStream(_input: string) { served: chunk.served_model || chunk.model || params.model, provider: chunk.provider_name || chunk.provider, requested: chunk.model || params.model, - fallback: !!(chunk.route && chunk.route.fallback_used), } } if (chunk.usage && !chunk.choices?.length) { diff --git a/frontend/src/views/Dashboard.vue b/frontend/src/views/Dashboard.vue index c671241..a2f8a13 100644 --- a/frontend/src/views/Dashboard.vue +++ b/frontend/src/views/Dashboard.vue @@ -41,16 +41,16 @@
- {{ t('dashboard.routeAlias') }} - {{ t('dashboard.routeChain') }} + {{ t('dashboard.byProtocol') }} + {{ t('dashboard.byProtocolSub') }}
-
+
{{ i + 1 }} {{ r.name }} {{ r.short }}
-
{{ t('dashboard.noData') }}
+
{{ t('dashboard.noData') }}
@@ -94,7 +94,6 @@ {{ t('dashboard.running') }} {{ t('dashboard.uptime') }} {{ uptimeText }} {{ t('dashboard.inflight') }} {{ health.in_flight ?? '-' }} / {{ health.in_flight_peak ?? '-' }} - {{ t('dashboard.fallback') }} {{ health.fallback_total ?? 0 }}
@@ -116,7 +115,7 @@ const { resolved } = useTheme() // 主题变化时重算图表配色 const health = ref({}) const metrics = ref({}) const models = ref({}) -const routes = ref([]) +const modelsList = ref([]) const loading = ref(false) const err = ref('') @@ -149,7 +148,6 @@ const cards = computed(() => { { k: t('dashboard.requests'), v: total.value, note: t('dashboard.noteOkFail', { ok: totalOk.value, fail: totalFail.value }) }, { k: t('dashboard.successRate'), v: rate.value.toFixed(1) + '%', note: t('dashboard.noteFail', { n: totalFail.value }) }, { k: t('dashboard.avgLatency'), v: avgP50.value.toFixed(2) + 's', note: t('dashboard.noteP95', { v: maxP95.value.toFixed(2) }) }, - { k: t('dashboard.fallbackShort'), v: health.value.fallback_total ?? 0, note: t('dashboard.noteInflight', { n: health.value.in_flight ?? 0 }) }, ] }) @@ -171,12 +169,14 @@ const providerRows = computed(() => { }) }) -const routeRows = computed(() => { - const m = models.value - return routes.value.map((r: any) => { - const chain: string[] = (m[r.id] || []).map((x: any) => (typeof x === 'string' ? x : x.id || x.name)) - return { name: r.id, short: chain.length ? t('models.chainCount', { n: chain.length }) : '—' } - }) +// 第二十七轮: 路由别名链已删除 → 这里改成"每条协议线各有几个模型" +// (三线直传后这个分布才是运维真正关心的:哪条线现在有货) +const protoRows = computed(() => { + const count: Record = {} + for (const m of modelsList.value) { + for (const p of (m.protocols || ['chat/completions'])) count[p] = (count[p] || 0) + 1 + } + return Object.keys(count).sort().map((k) => ({ name: k, short: t('dashboard.modelCountN', { n: count[k] }) })) }) const callRows = computed(() => { @@ -258,14 +258,10 @@ async function load() { try { const um = await api('/v1/models') const data = um.data || [] - routes.value = data.filter((x: any) => x.type === 'route') + modelsList.value = data const map: Record = {} - for (const x of data) if (x.type !== 'route') map[x.id] = x.provider_name || '' + for (const x of data) map[x.id] = x.provider_name || '' textModels.value = map - const list = data.filter((x: any) => x.type === 'route') - const m: Record = {} - for (const r of list) m[r.id] = r.models || [] - models.value = m } catch { /* 模型列表可选 */ } } catch (e: any) { if (e && e.status === 401) { router.push('/login'); return } diff --git a/frontend/src/views/GatewayConfig.vue b/frontend/src/views/GatewayConfig.vue index 7180907..e63daf8 100644 --- a/frontend/src/views/GatewayConfig.vue +++ b/frontend/src/views/GatewayConfig.vue @@ -122,7 +122,8 @@ async function load() { value: envs[env].value || '', show: false, })) - warns.value = (body.meta?.warnings || []).concat(body.meta?.route_warnings || []) + // 第二十七轮: 路由告警(meta.route_warnings)随路由功能删除,体检结论只看 warnings + warns.value = body.meta?.warnings || [] } catch (e: any) { err.value = (e?.body?.detail || e?.message || String(e)) } finally { diff --git a/frontend/src/views/Metrics.vue b/frontend/src/views/Metrics.vue index 876e591..beb8399 100644 --- a/frontend/src/views/Metrics.vue +++ b/frontend/src/views/Metrics.vue @@ -36,15 +36,14 @@
- {{ t('metrics.routeVsProvider') }} + {{ t('metrics.byProvider') }} {{ t('metrics.snapshot') }}
ok fail - fallback
- +
@@ -88,7 +87,6 @@
{{ t('metrics.snapshot') }} {{ t('metrics.uptime') }} {{ fmtUptime }} - {{ t('metrics.fallback') }} {{ data.fallback_total ?? 0 }} {{ t('metrics.inflight') }} {{ data.in_flight ?? 0 }} / {{ data.in_flight_peak ?? 0 }}
@@ -156,7 +154,8 @@ const cards = computed(() => { { k: t('metrics.requests'), v: all, note: t('dashboard.noteOkFail', { ok, fail }) }, { k: t('metrics.failCount'), v: fail, note: t('metrics.failRateNote', { v: rate.toFixed(1) }) }, { k: t('metrics.tokenUsage'), v: (tk.prompt || 0) + (tk.completion || 0), note: t('metrics.tokenNote', { p: tk.prompt ?? 0, c: tk.completion ?? 0 }) }, - { k: t('metrics.inflightPeak'), v: `${data.value.in_flight ?? 0} / ${data.value.in_flight_peak ?? 0}`, note: t('metrics.fallback') + ' ' + (data.value.fallback_total ?? 0) }, + // 第二十七轮: fallback 计数随路由功能删除 → 这张卡片的注脚改成运行时长 + { k: t('metrics.inflightPeak'), v: `${data.value.in_flight ?? 0} / ${data.value.in_flight_peak ?? 0}`, note: t('metrics.uptime') + ' ' + fmtUptime.value }, ] }) @@ -194,27 +193,26 @@ const provOpt = computed(() => { } }) -const routeOpt = computed(() => { +// 第二十七轮: 路由维度指标随路由功能删除(routes/fallback_total 已不在 /metrics 里), +// 这里改成按供应商看 成功 / 失败(供应商是现在唯一的"聚合维度")。 +const providerOpt = computed(() => { void resolved.value const c = chartTheme() - const r = data.value.routes || {} - const ks = Object.keys(r) - const names = ks.length ? ks : ['wjd'] - const get = (k: string, f: string) => (r[k] || {})[f] || 0 + const p = data.value.providers || {} + const names = Object.keys(p) + const get = (k: string, f: string) => (p[k] || {})[f] || 0 const ok = names.map((n) => get(n, 'ok')) const fail = names.map((n) => get(n, 'fail')) - const fallback = names.map((n) => get(n, 'fallback')) return { tooltip: { trigger: 'axis', axisPointer: { type: 'shadow' } }, - legend: legend(['ok', 'fail', 'fallback']), + legend: legend(['ok', 'fail']), grid, - graphic: emptyTip([ok, fail, fallback], t('metrics.none')), + graphic: emptyTip([ok, fail], t('metrics.none')), xAxis: categoryAxis(names, 10), yAxis: valueAxis('', 1), series: [ - { name: 'ok', type: 'bar', stack: 'r', data: ok, itemStyle: { color: c.a } }, - { name: 'fail', type: 'bar', stack: 'r', data: fail, itemStyle: { color: c.err } }, - { name: 'fallback', type: 'bar', stack: 'r', data: fallback, itemStyle: { color: c.warn } }, + { name: 'ok', type: 'bar', stack: 'p', data: ok, itemStyle: { color: c.a } }, + { name: 'fail', type: 'bar', stack: 'p', data: fail, itemStyle: { color: c.err } }, ], } }) diff --git a/frontend/src/views/Models.vue b/frontend/src/views/Models.vue index c510d50..7084387 100644 --- a/frontend/src/views/Models.vue +++ b/frontend/src/views/Models.vue @@ -11,36 +11,32 @@
- - +
- {{ t('models.routeAliasTitle') }} + {{ t('models.allTextModels') }} + {{ t('models.speedtestHint') }} + + + {{ latTesting ? t('models.testingShort') : t('models.speedtest') }} +
-
-
- {{ r.name }} - - - {{ latTesting ? t('models.testingShort') : t('models.speedtest') }} - -
-
-
{{ t('models.noRoute') }}
-
{{ t('models.allTextModels') }} {{ t('models.latency') }}
-
+
{{ m.id }} + {{ m.provider }} @@ -87,20 +83,42 @@ {{ t('models.modelListTitle') }} {{ t('models.capability') }}
- {{ t('models.routeCount', { n: routeRows.length }) }} + {{ t('models.textModelCount', { n: textModelRows.length }) }}
- + - + - + - + + + + + + - +