docs: 补 MIT LICENSE 与许可声明（G-Star 申报要求明确的许可证）

G-Star 合规要求下，把"模拟官方客户端、复刻私有协议、自动领取平台额度"这一类
能力整体移出仓库。这些实现依赖各平台未公开的端点、签名算法与官方 client id，
属于绕过客户端校验获取额度，风险高于普通协议适配。

移除（93 个文件）：
- 适配器：opencodezen（伪装 UA 取免费池）、trae、buddy、kimicode
- 插件：channels.lua（12 个平台登录/领额度/签到）、quota.lua（厂商额度探针）
- 随之删除的 78 个渠道/额度专用测试与 3 份渠道文档

同时修掉的既有问题：
- **硬编码第三方 OAuth client_secret**（Google Cloud Code 的 GOCSPX-…，两处）
  随 channels.lua 一起消失。
- `deploy.sh` 的 `sync_plugins` 整段删除：它仍按"内置插件是发行物"的旧前提
  把插件复制/覆盖进 plugin_dir，与"插件属运维资产、网关不安装不覆盖"的契约
  冲突。现在部署脚本只同步适配器与二进制/配置。
- 适配器清单去掉已移除项；e2e 配置改用显式 `auto:` 链（`models.priority` 早已
  移除，该测试此前已失效）。

保留（公开协议）：openai / anthropic / gemini / deepseek / github / groq /
mistral / ollama / qwen38 / sensenova = 10 个适配器；插件示例 billing / usage。

验证：go build -tags luajit ./... OK；internal/lua 21.4s、provider、core 51s、
gateway 122.8s、e2e 9.0s 全绿。

.snapshots/chat_nonstream.json

@@ -0,0 +1,24 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "message": {
+        "content": "pong",
+        "role": "assistant"
+      }
+    }
+  ],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion",
+  "usage": {
+    "completion": 1,
+    "completion_tokens": 1,
+    "prompt": 3,
+    "prompt_tokens": 3,
+    "total": 4,
+    "total_tokens": 4
+  }
+}

.snapshots/chat_stream.txt

@@ -0,0 +1,95 @@
+{
+  "choices": [
+    {
+      "delta": {
+        "content": "",
+        "role": "assistant"
+      },
+      "finish_reason": null,
+      "index": 0
+    }
+  ],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion.chunk"
+}
+{
+  "choices": [
+    {
+      "delta": {
+        "content": "Hel",
+        "role": "assistant"
+      },
+      "finish_reason": null,
+      "index": 0
+    }
+  ],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion.chunk"
+}
+{
+  "choices": [
+    {
+      "delta": {
+        "content": "lo",
+        "role": "assistant"
+      },
+      "finish_reason": null,
+      "index": 0
+    }
+  ],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion.chunk"
+}
+{
+  "choices": [
+    {
+      "delta": {
+        "content": "",
+        "role": "assistant"
+      },
+      "finish_reason": "stop",
+      "index": 0
+    }
+  ],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion.chunk"
+}
+{
+  "choices": [
+    {
+      "delta": {
+        "content": ""
+      },
+      "finish_reason": "stop",
+      "index": 0
+    }
+  ],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion.chunk"
+}
+{
+  "choices": [],
+  "created": <TS>,
+  "id": "<ID>",
+  "model": "good-m",
+  "object": "chat.completion.chunk",
+  "usage": {
+    "completion": 1,
+    "completion_tokens": 1,
+    "prompt": 16,
+    "prompt_tokens": 16,
+    "total": 17,
+    "total_tokens": 17
+  }
+}
+[DONE]

.snapshots/image_generation.json

@@ -0,0 +1,8 @@
+{
+  "created": <TS>,
+  "data": [
+    {
+      "b64_json": "QUJD"
+    }
+  ]
+}

HANDOFF_2026-10-19.md

@@ -0,0 +1,501 @@
+# ModelRouter / llmsproxy 交接文档
+
+> **历史文档**：本文记录的渠道能力（channels / quota 插件与 4 个私有协议适配器）
+> 已于 v3.0.2 按合规要求整体移出仓库。保留原文仅作历史记录，其中的路径、
+> 模块名与适配器清单不再代表当前代码。
+
+更新时间：2026-10-19
+工作目录：`/home/program/llmsproxy`
+当前分支：`release/v2.0.x`
+
+## 1. 当前 Git 状态
+
+最近提交：
+
+- `c9f3453 fix(codearts): 识别 InferHub 未注册模型`
+- `12bfd53 fix: 收口渠道源迁移与测试页合并`
+- `25970a2 refactor(scheduler)!: 移除 priority 废弃字段 + 一通道一源 + discover 复用凭据`
+
+`c9f3453` 已推送到 `origin/release/v2.0.x`。
+
+注意：当前工作树在最后一轮添加 provider 回归测试时出现了一个**未完成且可能无法编译的测试编辑**。最后一次工具报错：
+
+```text
+internal/provider/provider_test.go:99: missing ',' in argument list
+```
+
+下一位 agent 接手后的第一步必须是：
+
+```bash
+cd /home/program/llmsproxy
+git status --short
+gofmt -w internal/provider/provider_test.go 2>&1 || true
+go test -tags luajit ./internal/provider/ 2>&1
+```
+
+如果测试文件处于半编辑状态，先检查 `TestProviderChatStreamNormalizesStopAfterToolCall`，修正 Go 字符串转义或删除该未完成测试后再继续。不要直接部署当前工作树。
+
+## 2. 生产服务状态
+
+生产服务：`llmsproxy.service`
+
+- 当前服务 active
+- 监听：`0.0.0.0:8081`
+- 二进制已通过 `./deploy.sh` 原子替换并重启
+- 最近一次部署健康检查通过
+- 当前生产插件 SHA256 前缀曾核对为：`1bf26cbb3abcb107`
+- 当前生产二进制在最近部署后已更新
+
+不要把密钥写入交接记录。生产 admin key 保存在本机运维环境中，不应复制到聊天、文档或提交。
+
+## 3. Channel 源当前状态
+
+生产源已经从旧的“一账号一源”收口成“一渠道一源”：
+
+```text
+channels-trae
+channels-minimax
+channels-lobsterai
+channels-buddy
+channels-codearts
+channels-zcode
+```
+
+旧形态源已经清理：
+
+```text
+channels-<provider>-<account_id>
+```
+
+此前旧源和新源并存，造成同一渠道同一模型出现多份，只能靠账号 hex 区分。
+
+根因已修复：内核 `list_sources` 返回裸数组，但旧插件代码按 `{sources = [...]}` 信封读取，导致迁移清理循环永远为空。当前 `channels.lua` 已兼容：
+
+- 裸数组
+- `{sources = [...]}`
+- `{json = {sources = [...]}}`
+
+相关代码：
+
+```text
+internal/lua/plugins/channels.lua: syncProviderSource
+internal/core/channels_legacy_source_cleanup_test.go
+```
+
+当前生产源列表已实测只有 6 个新形态源。
+
+## 4. CodeArts 当前状态
+
+### 已确认可用的部分
+
+CodeArts 使用：
+
+```text
+adapter = codearts
+base_url = https://snap-access.cn-north-4.myhuaweicloud.com
+```
+
+CodeArts 的 chat 请求需要完整 AK/SK/security token 凭据，由插件适配器签名，不是裸 Bearer token。
+
+真实流式工具调用测试已经成功：
+
+请求形状：
+
+```json
+{
+  "model": "channels-codearts:openpangu-2.0-flash",
+  "messages": [{"role": "user", "content": "Use the weather tool for Beijing."}],
+  "tools": [{
+    "type": "function",
+    "function": {
+      "name": "get_weather",
+      "description": "Get weather",
+      "parameters": {
+        "type": "object",
+        "properties": {"city": {"type": "string"}},
+        "required": ["city"]
+      }
+    }
+  }],
+  "tool_choice": "auto",
+  "stream": true,
+  "max_tokens": 64
+}
+```
+
+结果：
+
+- HTTP 200
+- `Content-Type: text/event-stream`
+- 返回结构化 `tool_calls`
+- 函数名 `get_weather`
+- 参数最终为 `{"city":"Beijing"}`
+- `finish_reason = tool_calls`
+
+因此 CodeArts 的流式工具调用链路当前是通的。
+
+### CodeArts 模型刷新
+
+生产调用：
+
+```http
+POST /api/plugins/http/channels/model/refresh
+{"platform":"codearts"}
+```
+
+实测返回：
+
+```json
+{
+  "ok": true,
+  "results": [
+    {"platform": "codearts", "ok": true, "count": 11}
+  ]
+}
+```
+
+CodeArts 模型目录来自两个签名端点并合并：
+
+1. `opengw.developer.huaweicloud.com/api/v1/gateway/config`
+   - `result.models`
+2. `snap-access.cn-north-4.myhuaweicloud.com/v1/model/builtin`
+   - `builtinModels`
+
+相关解析：
+
+```text
+internal/lua/plugins/channels.lua: parseRemoteModels
+internal/lua/plugins/channels.lua: refreshRemoteModels
+internal/core/channels_codearts_catalog_test.go
+```
+
+### 已确认失效并禁用的 CodeArts 模型
+
+下面 4 个模型曾收到明确的上游 404：
+
+```text
+CodeArts-Pangu
+deepseek-v4-flash-0731
+deepseek-v4-pro-0813
+glm-5.3-flash
+```
+
+生产上已通过 `channels/model/set_disabled` 禁用。它们仍可能在 `model/list` 中出现，但 `disabled=true`，不会作为可用模型路由。
+
+新增的死模型判据：
+
+```lua
+model is not registered
+```
+
+对应提交：`c9f3453`。
+
+注意：当前插件的 dead-model 机制只在收到真实请求失败后记录模型；刷新目录不会自动根据“远端目录没有该模型”删除模型，这是刻意设计，避免误删远端目录不完整的平台模型。
+
+CodeArts 当前仍有一些模型返回 429/TPM 限流，例如：
+
+```text
+deepseek-v4-flash
+deepseek-v4.1-flash
+```
+
+429 不是模型失效，不能禁用。
+
+## 5. 其他 Channel 实测结果
+
+### Trae
+
+真实流式 + tools 请求：
+
+```text
+channels-trae:DeepSeek-V4-Flash
+```
+
+结果：
+
+- HTTP 200
+- SSE 正常返回
+- 结构化 tool call 返回
+- 但上游最后的结束原因可能是 `stop`，即使前面已经出现 tool-call fragments
+
+这会让 Agent 把工具调用误判成普通文本完成。
+
+当前准备中的修复在统一 provider 流式层：
+
+```text
+internal/provider/provider.go
+```
+
+逻辑是：只要本次流出现真实 tool-call 片段，后续 `FinishReason == "stop"` 就归一成 `tool_calls`。
+
+另外，Trae 的非流式接口偶尔返回 SSE。当前 `trae.lua` 对非流式 SSE 只取最后一条 data，不能完整合并多条工具调用/推理片段。这是后续仍需处理的协议问题，但不能用非流式手工请求结果推断 Pi 的真实请求模式。
+
+### Buddy
+
+真实流式 + tools 请求：
+
+```text
+channels-buddy:glm-5.3-flash
+```
+
+结果：
+
+- HTTP 200
+- SSE 正常返回
+- tool-call 参数分片可见
+- 最终有 `finish_reason = tool_calls`
+- 但某些上游 chunk 的 `tool_calls` 可能是空对象 `{}`
+
+空对象不是合法的 OpenAI tool-call 数组，部分 Agent 客户端会拒绝。
+
+当前准备中的修复在：
+
+```text
+internal/lua/adapters/buddy.lua
+```
+
+只在 `type(delta.tool_calls) == "table" and #delta.tool_calls > 0` 时输出 `tool_calls`，避免 Lua 空表被编码为 `{}`。
+
+Buddy 的非流式错误：
+
+```text
+11101: Non-stream chat request is currently not supported
+```
+
+之前出现的该错误来自手工 `curl` 测试中显式设置 `stream:false`，不能归因到 Pi。真实 `stream:true + tools` 已经证明 Buddy 流式工具调用可以工作。
+
+### MiniMax
+
+生产日志反复出现：
+
+```text
+adapter "anthropic": status=401 body={"error":"invalid access token"}
+provider channels-minimax: api error 401
+```
+
+这是凭据失效问题，不是 tools 形状问题。
+
+当前源：
+
+```text
+channels-minimax
+adapter = anthropic
+endpoint = /mavis/api/v1/llm/v1/messages
+```
+
+需要用户重新登录/提供有效凭据。不要擅自刷新一次性 refresh token，之前已经发生过 refresh token 被消耗后无法再次使用的事故。
+
+## 6. 当前未提交的代码改动
+
+最后一轮已经改动但尚未完成提交：
+
+```text
+internal/provider/provider.go
+internal/lua/adapters/trae.lua
+internal/lua/adapters/buddy.lua
+internal/provider/provider_test.go  # 可能有半编辑、需先修复
+```
+
+已完成的意图：
+
+1. provider 流式层记录 `sawToolCalls`
+2. 看到真实 tool-call 后，把后续 `stop` 归一为 `tool_calls`
+3. Trae 不输出空工具调用表
+4. Buddy 不输出空工具调用表
+5. provider 回归测试模拟“tool fragment + stop”场景
+
+但 provider 测试最后一次编辑工具报错：
+
+```text
+internal/provider/provider_test.go:99: missing ',' in argument list
+```
+
+因此必须先修测试文件，再提交。
+
+## 7. 推荐接手顺序
+
+### 第一步：修复工作树
+
+```bash
+cd /home/program/llmsproxy
+git status --short
+sed -n '1,150p' internal/provider/provider_test.go
+gofmt -w internal/provider/provider_test.go 2>&1 || true
+go test -tags luajit ./internal/provider/ 2>&1
+```
+
+若测试函数字符串转义难以修复，先删除半成品测试，再重新用 raw string 写测试。
+
+### 第二步：跑定向验证
+
+```bash
+luac -p internal/lua/adapters/trae.lua
+luac -p internal/lua/adapters/buddy.lua
+go test -tags luajit ./internal/provider/
+go test -tags luajit ./internal/lua/ -run 'Adapter|Buddy|Trae|Stream'
+```
+
+### 第三步：构建并部署临时验证版
+
+修复确认后：
+
+```bash
+go build ./...
+./deploy.sh
+```
+
+部署前确认：
+
+```bash
+sha256sum internal/lua/plugins/channels.lua
+sha256sum /etc/llmsproxy/plugins/channels.lua
+systemctl is-active llmsproxy
+```
+
+### 第四步：真实流式回归
+
+不要用 `stream:false` 测试流式问题。使用：
+
+```bash
+KEY='<从本机安全环境读取，不写入文档>'
+TOOL='{"type":"function","function":{"name":"get_weather","description":"Get weather","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}'
+
+curl --no-buffer --max-time 90 \
+  -H "Authorization: Bearer $KEY" \
+  -H 'Content-Type: application/json' \
+  -d "{\"model\":\"channels-codearts:openpangu-2.0-flash\",\"messages\":[{\"role\":\"user\",\"content\":\"Use the weather tool for Beijing.\"}],\"tools\":[$TOOL],\"tool_choice\":\"auto\",\"stream\":true,\"max_tokens\":64}" \
+  http://127.0.0.1:8081/v1/chat/completions
+```
+
+分别将 model 替换为：
+
+```text
+channels-codearts:openpangu-2.0-flash
+channels-trae:DeepSeek-V4-Flash
+channels-buddy:glm-5.3-flash
+```
+
+验收：
+
+- HTTP 200
+- `Content-Type: text/event-stream`
+- 不能出现 `tool_calls: {}`
+- 有 tool-call fragments 时最终必须是 `finish_reason: tool_calls`
+- CodeArts 当前应继续返回结构化工具调用
+
+### 第五步：CodeArts 刷新验收
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer $KEY" \
+  -H 'Content-Type: application/json' \
+  -d '{"platform":"codearts"}' \
+  http://127.0.0.1:8081/api/plugins/http/channels/model/refresh
+```
+
+预期：
+
+```json
+{"ok":true,"results":[{"platform":"codearts","ok":true,"count":11}]}
+```
+
+再检查：
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer $KEY" \
+  -H 'Content-Type: application/json' \
+  -d '{"platform":"codearts"}' \
+  http://127.0.0.1:8081/api/plugins/http/channels/model/list
+```
+
+预期：远端目录 11 个 + 兜底补位，已确认 404 的 4 个仍保持 `disabled=true`。
+
+## 8. 发布/安装包状态
+
+目前没有完成“全部安装包”发布闭环。
+
+已经发生的事情：
+
+- 核心包构建成功，但脚本实际使用了 `git describe`，产物版本类似：
+  `2.0.40~4-gc9f3453`
+- Linux GUI 包构建成功，版本为 `1.9.7`
+- Windows 主机 Wine 构建失败，生成了一个 264KB 的坏 `ModelRouter Setup 1.9.7.exe`，绝对不能发布
+- Docker Windows 构建也失败在拉取 Docker Hub 基础镜像 `golang:1.25` 网络超时
+- apt 无法找到 `wine32:i386`
+- 旧的 1.9.5/1.9.6 Windows 包是历史产物，不能作为当前版本
+
+不要把旧 Windows exe 当作新包发布。
+
+仓库已有 Docker 方案：
+
+```text
+cmd/gui/scripts/dist-win-docker.sh
+cmd/gui/docker/win-builder/Dockerfile
+```
+
+但当前 Docker Hub 网络不可用。若要继续做 Windows 包：
+
+1. 在可访问 Docker Hub 的构建机/CI 执行 Docker 方案；或
+2. 提供包含 Wine32/Wine64 的完整构建环境；或
+3. 暂时只发布 Linux/core 包，并明确说明 Windows 包未完成。
+
+版本脚本注意：
+
+- 核心包脚本使用位置参数：
+
+  ```bash
+  ./packaging/core-dist.sh 2.0.41
+  ```
+
+  不是 `VERSION=2.0.41 ./packaging/core-dist.sh`。
+- GUI 版本来自 `cmd/gui/package.json`，环境变量 `VERSION` 不会自动修改 Electron 产品版本。
+- 清理旧产物后再打包，并在全部包完成后统一计算 `SHA256SUMS`。
+
+## 9. 参考插件状态
+
+用户要求按：
+
+```text
+https://github.com/Ebony-Vinyl/dsh-our-free-model
+```
+
+逐厂商对照参考插件。
+
+当前本地路径：
+
+```text
+/tmp/pi-github-repos/Ebony-Vinyl/dsh-our-free-model
+```
+
+但该目录当前不是有效可读源码副本：
+
+- `vendor/channel-pack/src` 下面只有空目录；
+- `.git` 没有有效 `HEAD`/refs；
+- 无法从本地 Git 对象恢复源码；
+- 本地也没有 `buddy.ts`、`trae.ts`、`codearts.ts` 等文件。
+
+因此不能声称当前适配器是逐行照参考插件移植的。若要继续按用户要求完成，必须先重新 clone/恢复该 GitHub 仓库源码，确认以下文件实际存在后再对照：
+
+```text
+vendor/channel-pack/src/trae.ts
+vendor/channel-pack/src/buddy.ts
+vendor/channel-pack/src/codearts.ts
+vendor/channel-pack/src/*adapter*
+vendor/channel-pack/src/*model*
+```
+
+当前仓库代码只能作为临时协议修复基础，不能代替参考插件源码。
+
+## 10. 关键结论
+
+- Pi 是否发送流式请求，历史日志没有保存入站 `stream/tools` 字段，不能靠旧日志判断。
+- 本次真实 `stream:true + tools` 测试证明 ModelRouter 流式路径能工作：
+  - CodeArts：通过；
+  - Trae：通过但结束原因需归一；
+  - Buddy：通过但需过滤空 `{}`；
+  - MiniMax：凭据 401，独立问题。
+- 不要再用 `stream:false` 的手工请求推断 Pi 的问题。
+- CodeArts 刷新接口当前真实返回 `ok=true,count=11`，刷新链路不是当前阻塞点。
+- 当前最紧急的代码状态是修复 `provider_test.go` 的半编辑，然后验证未提交的统一流式修复。
+- 未经测试、未重新构建、未重新部署，不要继续改生产账号或刷新一次性凭据。

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 HeNanMU-EEA
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

Makefile

@@ -1,8 +1,8 @@
-.PHONY: all build test gui gui-dev gui-dist gui-deb gui-win gui-win-docker help clean
+.PHONY: all build test gui gui-dev gui-dist gui-deb gui-win gui-win-docker help clean gui-test
 
 # ModelRouter - single-binary gateway + optional Electron desktop GUI (embedded core)
 export BINARY
-BINARY=llmsproxy
+BINARY=modelrouter
 export BUILD_DIR
 BUILD_DIR=build
 export GUI_DIR
@@ -20,7 +20,7 @@ all: build
 # ---- core binary (server users / embedded core) ----
 build:
 	@mkdir -p $(BUILD_DIR)
-	CGO_ENABLED=1 go build -tags luajit -trimpath -o $(BUILD_DIR)/$(BINARY) ./cmd/llmsproxy
+	CGO_ENABLED=1 go build -tags luajit -trimpath -o $(BUILD_DIR)/$(BINARY) ./cmd/modelrouter
 	@echo "Built: $(BUILD_DIR)/$(BINARY) ($(VERSION))"
 
 # ---- core distribution packages (deb + rpm + tar.gz) -----
@@ -65,6 +65,13 @@ gui-dist-dir:
 test:
 	go test ./...
 
+# Desktop shell unit tests (no Electron needed): plugin-tree seeding/migration.
+# These guard the two things that silently break the desktop build:
+#   - plugin_dir missing from an embedded profile  -> channels plugin invisible
+#   - bundled plugin files never refreshed         -> old channels.lua survives upgrades
+gui-test:
+	@cd $(GUI_DIR) && node --test plugin-sync.test.js
+
 clean:
 	rm -rf $(BUILD_DIR) cmd/build cmd/gui/dist cmd/gui/build/gui-dist cmd/gui/bin
 
@@ -75,6 +82,7 @@ help:
 	@echo "  make gui        run desktop GUI (dev)"
 	@echo "  make gui-dev    run GUI with devtools"
 	@echo "  make gui-dist   build deb + rpm + AppImage (linux)"
+	@echo "  make gui-test   desktop shell unit tests (plugin tree seeding)"
 	@echo "  make gui-deb    build deb only"
 	@echo "  make gui-win    build Windows nsis (needs host mingw + wine)"
 	@echo "  make gui-win-docker  build Windows nsis all-in-docker (no host mingw needed)"

README.md

@@ -1,104 +1,51 @@
 # ModelRouter
 
+> **仓库**: <https://atomgit.com/HeNanMU-EEA/ModelRouter>
+>
 > **English**: [README_EN.md](./README_EN.md)
 >
 > **AI-assisted**: 本项目使用 AI 辅助编程（代码与文档由 AI 协作完成，经人工复核）。
 
-面向多 llms 订阅者，部署在内网，实现一次配置多个服务共同使用的效果。支持指定模型或 auto 模式，按照配置的优先级选择可用模型提供服务。
+面向多 LLM 订阅者的内网网关：一次配置，多个服务共用一组上游额度。支持指定模型或
+`AUTO` 模式，按配置的优先级选择可用模型提供服务。
 
-> **轻量且增源无需重编译**：网关本体是单 Go 二进制（约 10MB，零运行时依赖）。新增/切换上游
+网关本体是单个 Go 二进制（约 10 MB，零运行时依赖）。新增或切换上游只需在
-> 只需在 `config.yaml`（或 WebUI）加一个 `sources` 条目或挂一个 `.lua` 适配器——**不改 Go、
+`config.yaml`（或 WebUI）加一个 `sources` 条目、或挂一个 `.lua` 适配器——**不改 Go、
-> 不重编译**。源与适配器的改动经 WebUI 提交时即时生效（热更新）；直接编辑
+不重编译**。源与适配器的改动经 WebUI 提交时即时生效（热更新）；直接编辑
-> `config.yaml` 或 `adapter_dir` 下的 `.lua` 文件则需要重启进程生效。
+`config.yaml` 或 `adapter_dir` 下的 `.lua` 文件需要重启进程生效。
 
-## 核心优势
+## 核心能力
 
-### 极致轻量
+- **统一接口**：对内暴露 OpenAI 兼容的 `/v1/chat/completions`、`/v1/models`、
-
+  `/v1/images/generations`；所有源对外都是同一套格式（含 `reasoning_content`、
-- **单二进制**：编译后 8~12MB（`-s -w` strip 后约 8MB），零运行时依赖（仅依赖系统 libc），部署即用
+  `tool_calls`、`usage`）。
-- **低内存占用**：与源数量相关，非与运行时长相关——单源 ~10MB、16 源生产实例 ~32–35MB（[实测分解与调优](#内存占用实测与调优)）
+- **多源 + AUTO 链**：一个进程内任意多个上游源；`default_model: AUTO` 时按优先级页
-- **日志按需加载**：统计数字来自**全量**审计日志（启动时流式扫一遍即释放，29MB/22 万行约 260ms），但原始记录不常驻内存——默认只加载首屏，下滚自动分页，CSV 导出流式写出（O(1) 内存），离页即释放
+  保存的档位逐档调度，源故障自动降级并冷却，恢复后半冷却探测自愈。
-- **零运行时依赖**：纯 Go + LuaJIT 静态链接，无需安装 Python/Node/Java 等运行时
+- **协议适配在 Lua 层**：每个源挂一个 `.lua` 适配器，完成请求/响应/流式分片的双向
-- **启动极快**：冷启动 < 200ms，热重载配置 < 10ms
+  转换，并可用 `build_headers(meta)` 注入签名头。协议差异不进 Go 内核。
-
+- **插件独立**：业务扩展（计费、用量统计等）是外置 Lua 插件，由插件目录加载，
-### 强大的多租户调度能力
+  不编进二进制、不随发行包分发；插件只通过 `modelrouter` SDK 与内核交互。
-
+- **多租户密钥**：网关 key 与上游 key 相互独立，按 key 限定模型范围、AUTO 链与配额。
-- **多密钥多租户**：支持无限密钥，每个密钥独立角色、模型范围、Token 配额、重置周期
+- **加密存储**：上游 key、自定义请求头、网关 key 均以 AES-256-GCM 落盘，
-- **按模型配额**：每把 key 的每个模型单独配 token + 请求数配额与重置周期（小时/周/月/自定义 N 小时）；一个模型用满只影响该模型，同 key 其它模型照常；耗尽返 429 + `Retry-After` 并点名模型，admin key 永不受限
+  `master.key` 单独 0600 权限（或由 `MODELROUTER_MASTER_KEY` 注入）。
-- **AUTO 智能调度**：基于优先级档位的分级调度，同优先级源自动轮询负载均衡，故障自动毫秒级故障转移
+- **可观测**：状态页实时探测源可达性；HTTP 访问日志、配置变更审计、请求与用量 CSV 导出；
-- **自愈冷却**：冷却上限 5 分钟，过半后放行 1 个探测请求，上游/额度恢复即刻回归轮询，无需等满冷却窗口
+  写操作不落在请求热路径上。
-- **Token 配额管理**：精确到模型级别的 Token 配额控制，支持小时/周/月/自定义小时周期自动重置
-- **同优先级源负载均衡**：同一优先级档位的多个源，请求自动 Round-Robin 均匀分发，故障毫秒级故障转移
-- **Source-Model 前缀路由**：支持 `source-model`/`source:model`/`source/model` 精确指定上游源
-
-### 生产级可靠性
-
-- **热加载配置**：WebUI 修改源/密钥/适配器/AUTO链即时生效，无需重启
-- **密钥加密落盘**：AES-256-GCM 加密存储 `api_key`、网关密钥、请求头，master.key 0600 权限保护
-- **实时源探测**：`GET {base}/models` 定期探测，毫秒级感知上游状态，不污染调度退避状态
-- **全链路审计**：HTTP 访问日志、登录/配置变更审计、请求记录 CSV 导出、密钥用量统计 CSV 导出
-
-### 灵活的协议适配
-
-- **LuaJIT VM**：每个适配器独立 VM + **弹性 worker 池**（按需扩缩容，空闲时自动回收），安全并发，Lua 脚本热加载无需重启
-- **Lua 适配器协议**：`transform_request` / `transform_response` / `transform_stream_chunk` 双向转换
-- **动态请求头钩子**：`build_headers(meta)` 支持 HMAC 签名、动态 Header 注入
-- **多模态透传**：`image_url` 等多模态内容在多源间无损透传，Anthropic/Gemini/Ollama 自动转换
-
-## 实现
-
-统一 OpenAI 兼容网关：把多个上游 LLM 源（DeepSeek、Qijiar、OpenAI、Anthropic、Gemini、
-Groq、Mistral、Ollama、KimiCode…）通过 **Lua 适配器** 做协议转换，对内网暴露
-一个标准 `OpenAI Chat Completions` 接口（`/v1/chat/completions` + `/v1/models`），
-支持单次结算、SSE 流式、**AUTO 模型路由**与**多模态**透传。
-
-从 [HomeAgent](https://gitcode.com/JianFeeeee/HomeAgent) 的多源 LLM 适配层
-（`internal/agent/api/provider.go` + `internal/lua/adapters/*`）抽离而来并独立演进。
-
-## 特性
-
-- **多源**：一个进程内配置任意多个上游源，按请求的 `model` 自动路由。
-- **AUTO 模式**：`default_model: AUTO` 时按优先级页保存的 AUTO 链档位逐档调度并发请求。
-- **统一输出**：所有源对外都是 OpenAI 格式（含 `reasoning_content`、`tool_calls`、`usage`）。
-- **生图**：`POST /v1/images/generations`，`kind: image` 的模型独立路由。
-- **多模态**：`content` 数组（`image_url` 等）在多源间无损透传，Anthropic/Gemini/Ollama 自动转换。
-- **LuaJIT VM**：基于 golua 绑定的 LuaJIT，每个适配器独立 VM + worker 池，安全并发。
-  池是 **弹性** 的：`Σ max_concurrent` 只是上限而非预分配，状态按需创建、空闲时自动回收，
-  因此空载网关几乎不持有 Lua 状态（启动 0 个）。
-  - 扩容步长由适配器最大并发决定：`clamp(ceil(max/8), 1, 8)`，且仅在**真实并发争用**
-    （所有现有状态均已占用）时批量预热，顺序流量始终只用 1 个状态。
-  - 缩容步长由当前连接数决定：`clamp(ceil(冗余/(1+占用)), 1, 冗余)`——无连接时一轮收到
-    常驻下限，忙时每轮只释放 1 个，保护热路径。WebUI「适配器」页展示实时池指标。
-- **disable_thinking**：请求 `disable_thinking:true`（或上游对应字段）动态开关推理。
-- **Lua 适配协议**：每个源挂一个 `.lua` 适配器，完成 `transform_request` /
-  `transform_response` / `transform_stream_chunk` 双向转换，协议差异全在 Lua 层。
-- **签名 / 请求头钩子**：适配器可定义 `build_headers(meta)`，在 Go 发 HTTP 前
-  动态注入/签名请求头——用于云端 API 校验调用方 app（如 KimiCode 只放行特定
-  agent）。提供 `hmac_sha256_hex` / `sha256_hex` / `base64_encode` 等签名辅助。
-- **WebUI**：内置管理界面，可在线查看/新增/编辑上游源与模型、配置 AUTO 优先级链、
-  管理密钥模型范围，写入运行时文件持久化。
-- **密钥加密存储**：运行时文件中的上游 `api_key`、自定义请求头值、网关 key 均以
-  AES-256-GCM 加密落盘（`master.key` 独立 0600，或 `LLMS_PROXY_MASTER_KEY`）。
-- **实时源探测**：状态页探测各源可达性（`GET {base}/models`，失败回退最小请求），
-  不污染正常调度的退避状态，错误信息在 UI 可悬停查看。
-- **鉴权**：网关自身用 `gateway_keys` 校验客户端 Bearer key；与上游各自的 key 相互独立。
-- **流式**：SSE `chat.completion.chunk`，含角色首包与 `[DONE]` 收尾。
-- **全链路审计**：HTTP 访问日志、登录/配置变更审计、请求记录 CSV 导出、密钥用量统计 CSV 导出
-- **AES-256-GCM 加密存储**：运行时文件敏感字段加密落盘，master.key 0600 权限，支持环境变量注入主密钥
 
 ## 快速开始
 
 ```bash
-GOMODCACHE=... GOPROXY=off go build -tags luajit -o llmsproxy ./cmd/llmsproxy
+# 需要 LuaJIT 开发库；-tags luajit 是必须的
-./llmsproxy -config config.yaml    # 首次运行自动生成默认配置并打印随机 admin key
+GOMODCACHE=... GOPROXY=off go build -tags luajit -o modelrouter ./cmd/modelrouter
+
+# 首次运行会生成默认配置，并在日志里打印随机 admin key
+./modelrouter -config config.yaml
 ```
 
-> 依赖 [golua](https://github.com/aarzilli/golua)（LuaJIT 绑定）。**必须**带 `-tags luajit`
+> 依赖 [golua](https://github.com/aarzilli/golua)（LuaJIT 绑定）。不带 `-tags luajit`
-> 构建，否则默认走内置超集 gopher-lua 路径（行为略有差异）。
+> 会走 gopher-lua 路径，行为略有差异。
 
-> **配置不入库**：仓库不携带任何 config 文件（配置文件含密钥）。二进制首次运行时
+> **配置不入库**：仓库不携带任何 config 文件（其中含密钥）。默认配置只绑定
-> 会在 `-config` 指定的位置生成默认配置：随机 admin key（打印在启动日志中）、
+> `127.0.0.1:8080`，首次登录后请在 WebUI「密钥」页更换管理员密钥。
-> 仅绑定 `127.0.0.1:8080`。首次登录后请在 WebUI「密钥」页更换管理员密钥。
 
 ```bash
 # 无 key -> 401
@@ -115,474 +62,77 @@ curl -N -H "Authorization: Bearer $KEY" \
      http://127.0.0.1:8080/v1/chat/completions
 ```
 
-任何 OpenAI SDK 把 `base_url` 指到网关地址、`api_key` 用任意一个已创建并授权的网关 key 即可。
+任何 OpenAI SDK 把 `base_url` 指向网关、`api_key` 换成任一已授权的网关 key 即可。
 
-## 配置
+## 文档
 
-首次运行在 `-config` 路径自动生成默认配置（默认 `./config.yaml`），核心字段：
+| 主题 | 文档 |
+|---|---|
+| 实现与完整特性清单 | [docs/features.md](docs/features.md) |
+| 配置：`config.yaml`、自动探测模型、多密钥 | [docs/configuration.md](docs/configuration.md) |
+| 路由：模型路由、AUTO 链、冷却与半冷却、源与适配器、源模板 | [docs/routing.md](docs/routing.md) |
+| 适配器：Lua 协议、内置适配器清单 | [docs/adapters.md](docs/adapters.md) · [docs/lua-adapters.md](docs/lua-adapters.md)（[EN](docs/lua-adapters-en.md)） |
+| 插件：SDK 契约、钩子、UI 扩展、HTTP 端点 | [docs/plugins.md](docs/plugins.md) |
+| 桌面端：Headless / Desktop 两种产物、构建与运行 | [docs/gui.md](docs/gui.md) |
+| 性能：实测内存占用与调优、thinking 预算 | [docs/performance.md](docs/performance.md) |
+| WebUI 使用 | [docs/webui.md](docs/webui.md) |
+| 开发：目录职责、测试、分支与发布纪律 | [docs/development.md](docs/development.md) · [docs/git-workflow.md](docs/git-workflow.md)（[EN](docs/git-workflow-en.md)） |
+| 部署与迁移 | [docs/deployment.md](docs/deployment.md) |
 
-```yaml
+## 从 llmsproxy 改名
-listen: 127.0.0.1:8080        # 网关监听地址（建议绑内网/回环）
-gateway_keys: [sk-gw-<随机>]  # 初始 admin 密钥（每次安装随机生成，仅首启写入运行时存储用）
-default_model: AUTO           # model 无法路由时自动按优先级链选源
-adapter_dir: adapters         # Lua 适配目录；目录不存在时首启创建并 seed 内置，存在则只读
-runtime_file: runtime.json    # WebUI 编辑的源/密钥/AUTO 链持久化到此文件
 
-sources:
+产品名早已是 ModelRouter，二进制、配置目录、服务名与环境变量现已统一为
-  - name: deepseek
+`modelrouter`：
-    base_url: https://api.deepseek.com
-    api_key: sk-...            # 也可用 api_key_env: SOME_ENV 引用环境变量（不落盘明文）
-    adapter: deepseek
-    max_concurrent: 8
-    models:
-      - id: deepseek-v4-flash
-        priority: 100          # YAML 源首次启动会 seed 进 AUTO 链的初值
-        kind: chat
-      - id: deepseek-v4-pro
-        priority: 60
-        kind: chat
-    # 静态请求头（比适配器默认优先）
-    headers: { X-Tenant: prod }
-    # 传给 Lua build_headers 的透传元数据
-    meta: { app_id: x, app_secret: y }
-    temperature: 0.7
-    max_tokens: 4096
-    timeout: 120s              # 请求超时，默认 120s
-```
 
-`config.yaml`、`runtime.json`、`master.key`、`adapters/` 均不提交版本库（见
+| 旧 | 新 |
-`.gitignore`），请随备份带走——`master.key` 丢失后运行时密文无法解密。
+|---|---|
+| `llmsproxy`（二进制） | `modelrouter` |
+| `/etc/llmsproxy/` | `/etc/modelrouter/` |
+| `llmsproxy.service` | `modelrouter.service` |
+| `LLMSPROXY_*` | `MODELROUTER_*` |
 
-运行文件生成在 `-config` 所在目录：`adapter_dir` / `runtime_file` 自动指向
+已部署的机器用迁移脚本升级（**默认 dry-run**，复制不移动，可回滚）：
-配置文件旁边的 `adapters/` 与 `runtime.json`，从任意工作目录启动都可用。
-
-运行时文件（`runtime_file`）中的敏感字段自动加密：
-
-- 加密算法 AES-256-GCM，格式 `enc:v1:<base64>`。
-- 主密钥来源：环境变量 `LLMS_PROXY_MASTER_KEY`（64 位 hex）；否则读取
-  `runtime_file` 同目录的 `master.key`；都不存在时首次启动自动生成 `master.key`（0600）。
-- 升级时旧明文文件自动兼容：首次运行正常读取，任何 UI 保存操作触发全文件加密迁移。
-- 注意：**master.key 丢失后密文无法解密**，请随配置一起备份；切勿提交到版本库。
-
-### 网关密钥（多密钥）
-
-网关鉴权采用「多密钥 + 角色 + 模型范围」架构，密钥持久化在 `runtime_file` 的
-`keys` 字段（加密存储）：
-
-- `gateway_keys` 配置只是**初始 admin 密钥种子**：首次启动迁移为运行时 admin
-  key，之后不再参与鉴权管理。
-- **重要：首次启动后请在 WebUI「密钥」页更换管理员密钥**——初始密钥明文写在
-  `config.yaml` 里（每次安装随机生成），继续使用存在被盗风险；用新密钥登录后删除初始密钥。
-- WebUI **密钥页**可创建/删除密钥；每个密钥可指定 `admin`（管理全部）或
-  `user`（仅看自己的 key）角色，并配置**模型范围**（模型 + 源 + token 配额 +
-  重置周期）。
-- 客户端用任意一个已授权的密钥明文作为 Bearer（`Authorization: Bearer <key>`）。
-- 删除密钥即从运行时存储移除，立即失效。
-
-#### 密钥用量配额
-
-**配额按模型单独设置**：每个密钥的 `models[]` 里，每一条模型范围各自带一份
-token 配额与请求数配额。一个模型用满只影响该模型，同一密钥的其它模型照常工作。
-
-```yaml
-keys:
-  - key: sk-gw-<hex>
-    role: user
-    name: agent-alice
-    models:
-      - model: deepseek-v4-flash
-        token_quota: 1000000   # 本周期内该 key 用这个模型的 token 预算
-        req_quota: 20000       # 本周期内的请求次数
-        period: nhour          # "" | hour | week | month | nhour
-        hours: 6               # 仅 nhour
-      - model: AUTO
-        token_quota: 5000000   # AUTO 也是一条独立配额
-        period: day            # ← 这种写法会被拒绝（词表只有 hour/week/month/nhour）
-      - model: kimi-k3        # 未列配额 = 无限
-```
-
-- **没有「整钥总配额」**：这是刻意的设计。整钥总额会让「把 A 模型的额度挪给
-  B 模型」变成一次全局重分配；按模型独立计费则每个模型各自可控，运维可以
-  看出哪个模型吃掉了预算。
-- `period` 词表：空 = 永不过期（累计总量），`hour` / `week` / `month` = 固定窗口，
-  `nhour` + `hours` = 自定义小时数。**拼错的周期在写入时就被拒**，不会静默变成
-  永不过期。
-- 配额严格按密钥隔离，且**同一密钥内按模型隔离**：A 密钥用满 `m1` 不会消耗
-  B 密钥的额度，同一密钥的 `m2` 也不受影响。
-- 配额统计含聊天、流式、生图，跨重启从审计日志回放（保留 40 天，覆盖最长的
-  month 窗口）。
-- 配额耗尽返回 **429 + `Retry-After`**（`rate_limit_exceeded`），消息里点名是哪个
-  模型用满了，客户端可等窗口重置后自动恢复；模型越权才是 403。
-  **admin 密钥永不受配额限制**（其 scope 上的配额也不强制），避免把管理员
-  锁在门外。
-- 窗口用量按整点小时分桶统计，实际释放比配置窗口最多晚 1 小时（配额宁可晚释放
-  也不超发）。
-- `PUT /api/keys/{key}` 提交 `models` 即同时提交它们的配额（配额就是 scope 的一部分，
-  不存在会与模型列表脱节的第二份预算）。显式 `0` = 解除该模型的限制。
-
-##### 配额拒绝 vs 容量拒绝：两种「拒绝」含义不同
-
-配额在**入口**检查，不占用任何上游槽位，因此是廉价拒绝；容量不足则要走
-`busyWait` 有界等待后才返回 503。实测（容量 8、0.6s/请求、100 并发）：
-
-| 情况 | 状态码 | 延迟 | 是否打到上游 |
-|---|---|---|---|
-| 配额用尽（`req_quota` 50 → 100 并发） | `429 rate_limit_exceeded` + `Retry-After` | **19–21ms** | 否 |
-| 容量打满（同一拓扑，未限配额） | `503 upstream_error` | ~2.6s（`busyWait`） | 超出部分否 |
-
-客户端因此可以区分「等窗口重置」与「等上游腾出容量」：前者只需等待，后者
-通常该降并发或换源。
-
-##### 配额的开销
-
-配额桶按 (密钥, 模型, 整点小时) 分桶保留 40 天，实测（AMD 7840HS）：
-
-- 窗口查询按窗口长度而非保留总量扫描：1h 窗口 49ns、24h 窗口 55ns、
-  30d 窗口 3.9μs（此前全扫保留总量，960 桶时 5.9μs）。
-- 记录一条请求：283ns、3 分配（与引入配额前相同，分配来自 ring buffer）。
-- 内存：生产形态（7 密钥 × 8 模型 × 2 源 × 满 40 天 retention）约 **3.7MB**。
-  按源 pin 的 `source::model` 桶**惰性创建**——只有当某条配额真的 pin 了
-  某个源时才维护，否则每条记录多写一份桶，在 20 密钥 × 8 模型 × 3 源下会
-  多占 18MB。
-  代价：配置 pinned 配额之前的历史用量无法事后按源拆分，所以 **pinned 配额的
-  首个窗口可能少算**。
 
 ```bash
-# 配额耗尽时客户端看到（点名了具体模型）
+sudo ./packaging/migrate-to-modelrouter.sh              # 预演
-HTTP/1.1 429 Too Many Requests
+sudo ./packaging/migrate-to-modelrouter.sh --apply      # 执行
-Retry-After: 2100
+sudo ./packaging/migrate-to-modelrouter.sh --rollback --apply
-{"error":{"type":"rate_limit_exceeded","message":"token quota exceeded for \"deepseek-v4-flash\" (5000000/5000000)"}}
 ```
 
-### 模型路由
+旧路径与旧环境变量在读取侧仍被接受，因此升级不会因为顺序问题立刻失效。
-
+`deploy.sh` 在检测到旧 `/etc/llmsproxy` 而新配置不存在时会**中止**，避免部署出
-`/v1/chat/completions` 的 `model` 解析顺序：
+一个「能启动但没有 key / 没有源」的空实例。
-
+
-1. `source/model` 或 `source:model` 前缀 → 指定源；
+插件 SDK 同名调整：唯一公开入口是 `modelrouter`（`llmsproxy` 仅作兼容别名），
-2. 精确匹配某个源的 `model`；
+详见 [docs/plugins.md](docs/plugins.md)。
-3. 配置 `default_model: AUTO` 时 → 走优先级页保存的 AUTO 链（见下）；
+
-4. 否则回落（返回错误）。
+## 目录结构
-
+
-任何 OpenAI 客户端，只要 `model` 设为某个源的 `name/任意名`，即可锁定走该源；
+```text
-设为 `AUTO`（或网关配了 `default_model: AUTO`）即自动按优先级链选源。
+cmd/modelrouter          # 入口（纯 Go 服务端）
-
+cmd/gui                  # Electron 桌面壳（内嵌同一份核心）
-### AUTO 链（优先级页）
+internal/config          # 配置解析、加密存储、运行时文件
-
+internal/core            # 调度、内核动作、状态与审计
-AUTO 调度**只**由优先级页保存的规则决定，源配置里的 `models[].priority` 数字不再
+internal/lua             # LuaJIT VM + 弹性 worker 池 + 内置适配器 (embed)
-参与调度、也不再显示。**聊天与生图是两套独立的链**，在优先级页面通过「聊天 / 生图」
+internal/lua/plugins     # 可选插件示例（不随二进制/发行包分发）
-按钮切换编辑：
-
-- **聊天链**（`runtime_file` 的 `auto` 字段）：服务 `POST /v1/chat/completions`
-  的 `model: AUTO` 请求；只接受 `kind: chat` 模型，配置在链里的生图槽位会被忽略。
-- **生图链**（`runtime_file` 的 `auto_image` 字段）：服务 `POST /v1/images/generations`
-  的 `model: AUTO` 请求；只接受 `kind: image` 模型。未配置生图链时，AUTO 生图回落
-  到「所有含生图模型的源、按注册顺序」的旧行为，兼容旧配置。
-
-每条规则 = 一个「槽位」：`{ model, source, tier, token_quota, period, hours }`。
-
-- `tier` 表示优先级档位：同一档的模型并排、共享该优先级；档位从上到下递减。
-- 同一模型可配置多个槽位（如 A 源低配 → B 源低配 → A 源高配 → B 源高配），按档位顺延。
-- `token_quota` > 0 时该槽位在重置周期内用满即顺延到下一槽位；`period` 支持
-  `hour` / `week` / `month` / `nhour`（配合 `hours`），空 = 不限。
-- 同 tier 内按「偏好分」排序，失败槽位冷却后自动跳过；冷却 / 配额耗尽 / hard
-  错误会顺延到下一 tier，全部失败时返回 503 并附上每档失败摘要。
-
-### 冷却与半冷却探测（自愈调度）
-
-槽位失败后进入指数退避冷却（5s → 10s → 20s …，上限 **5 分钟**）。冷却不是全时段
-硬闸：
-
-- **前半段完全静默**：刚失败的槽位不接任何流量，避免对故障上游持续施压。
-- **后半段放行 1 个探测**：窗口过半后，同一 `(源, 模型)` 最多允许 **一个** 请求作为
-  探测通过。探测槽数 = `clamp(max_concurrent / 10, 1, 2)`——`max_concurrent: 10`
-  的源恰好放行 1 个探测请求。
-- **探测排在最后**：同 tier 内探测候选永远排在健康候选之后，只有没有正常槽可用时
-  才承接真实流量，因此探测不会抢走可用容量。
-- **探测即真实请求**：成功则立即清零失败计数与冷却，槽位当场回到正常轮询——
-  上游恢复/额度恢复后不必等满整个冷却窗口。失败则重开一个新窗口，下次探测顺延到
-  新窗口的中点，不会连续重试。
-
-分类冷却策略：
-
-| 情况 | 冷却 | 说明 |
-|---|---|---|
-| 传输错误 / 5xx | 5s→10s→20s…，上限 5min | 指数退避 |
-| 401 / 403 凭据错误 | 10min | 换 key 后由下一次探测自动接回 |
-| 429 限流 | 30s 固定 | 「太快」不等于「坏了」，不进指数阶梯 |
-| 配额耗尽（`insufficient_quota` / `余额不足` 等） | 对齐配额窗口，上限 30min | **不进指数阶梯、不涨失败计数**，额度恢复即可用 |
-
-WebUI「优先级」页的槽位健康标签会区分 `冷却` / `待探测` / `探测中`，悬停可看窗口
-起点、探测放行时刻与完全恢复时刻。
-
-### 源（Source）与适配器（Adapter）的关系
-
-- **源** 是到某个上游的连接描述：`name`、`base_url`、`api_key`、模型列表与优先级。
-- **适配器** 是协议转换逻辑（Lua 脚本）：把统一 OpenAI 格式请求转成上游原生格式，
-  再把上游响应/流式分块转回统一格式。
-- 一个适配器可被多个源复用（如 `openai.lua` 同时服务多个 OpenAI 兼容站点）；
-  同一个源也可以切换不同适配器（改 `adapter` 字段即可）。
-- 源决定“连谁、暴露哪些模型”，适配器决定“怎么对话”——二者在 `sources[]` 条目中
-  通过 `adapter` 字段关联。
-
-加载流程（`internal/core` 的 `Core` 负责装配）：
-
-1. 启动时 `lua.NewVM(adapter_dir)` 加载全部适配器：内置适配器（编译期 embed）+
-   `adapter_dir` 下的同名覆盖文件；内置文件写在代码内，覆盖文件要求更高优先级。
-2. `config.Load` 读取 `config.yaml`，`config.NewStore(runtime_file)` 读 WebUI 改动的
-   运行时源，二者按名称合并成完整源列表。
-3. `rebuildRegistry` 为每个源创建 `provider.Provider`（持有目标适配器），并按源的
-   并发上限配置适配器 worker 池**上限**（不预分配，见下）。
-4. 请求进来时 `Registry` 按 `model` 路由到 Provider，Provider 调适配器
-   `transform_request` → HTTP 发送 → `transform_response` / `transform_stream_chunk`。
-
-WebUI 上的"新增/编辑源"、"上传 Lua 适配器"、"改 AUTO 优先级链"与"密钥模型范围"
-都即时生效（写入运行时文件或 `adapter_dir` 后重新装配，无需重启）；直接编辑
-`config.yaml` / `adapter_dir` 下的文件则需要重启进程才会重新加载。
-
-在 WebUI 删除源/适配器时：运行时（WebUI 创建）的源与 `.lua` 文件会真正移除；
-`config.yaml` 中定义的基源无法改写配置文件，采用删除标记隐藏（重启后仍隐藏），
-在 UI 里重新添加同名源即可恢复。
-
-### 源模板（多 key 负载均衡）
-
-多个上游 key 共享同一套配置（URL / 适配器 / 模型列表 / 并发 / rpm）时，逐一复制
-整段 source 太冗余。**源模板**保存除 `name` 和 `api_key` 以外的全部源字段，
-从模板一键创建多个仅 key 不同的源，天然支持多 key 负载均衡（每个展开源独立调度、
-独立冷却、独立健康状态）。
-
-- WebUI「源」页右上角「**模板管理**」按钮：列出全部模板，可编辑 / 删除 / 新建。
-- 添加源弹窗右上角两个按钮：
-  - **从模板创建** → 选模板 → 表单自动填充（name + key 仍需你填）；
-  - **存为模板** → 把当前表单非 key/name 字段存为新模板。
-- 模板持久化在 `runtime_file` 的 `source_templates` 字段，与运行时源一样热管理，
-  改完即时生效。模板本身不含密钥，可以安全共享 / 纳管。
-
-模板不会自己变成源——它只是「配方」。真正承担流量的是从模板创建出来的、带 key
-的具体源。
-
-### 内存占用（实测与调优）
-
-内存不是一个定数：它随**配置的源数量**增长（每个源一个 `http.Transport` 与连接池、
-一组健康状态），与运行时长无关。本机实测（Linux x86_64，12 核）：
-
-| 部署形态 | 启动 RSS | 稳态 RSS |
-|---|---|---|
-| 1 源 / 1 适配器（最小配置） | ~4 MB | ~10 MB |
-| 1 源 + 29 MB 历史审计日志 | ~19 MB | ~20 MB |
-| **16 源 / 13 适配器 / 59 模型（本机生产）** | ~28 MB | **~32–35 MB** |
-
-> 历史参考：本项优化前同一生产配置为 **~105 MB**。降幅来自三处：审计日志的**原始记录**
-> 不再常驻内存（约 25 MB；统计聚合仍扫全量，但扫完即释放）、Lua 状态池不再单调增长、
-> 以及下面两个运行时开关。
-
-内存构成（生产实例分段测量，`/proc/<pid>/smaps`）：
-
-| 区域 | RSS | 说明 |
-|---|---|---|
-| Go 堆 | ~14 MB | provider/registry/scheduler 结构 + 连接池缓冲 |
-| 其他匹名（线程栈 / LuaJIT chunk / runtime） | ~12 MB | 与线程数、已加载适配器数相关 |
-| 二进制 text+rodata | ~8 MB | 映射的可执行文件页（只读、可被内核回收） |
-| Go runtime 预留 | ~4 MB | VSZ 上看到的 2 GB+ 是地址空间预留，不占物理内存 |
-| 共享库 | ~3 MB | libc / libluajit / libm |
-
-**两个推荐的部署开关**（只适用于环境变量，无需改代码）：
-
-```ini
-# /etc/systemd/system/llmsproxy.service
-Environment=GOGC=50
-Environment=MALLOC_ARENA_MAX=2
-```
-
-- `MALLOC_ARENA_MAX=2`：LuaJIT 的分配走 cgo → glibc malloc，glibc 默认允许 `8×nproc`
-  个 per-thread arena，每个碰到 malloc 的 OS 线程会占用一个（各约 1 MB，且**不归还给系统**）。
-  实测从 8–15 个 arena（约 7–12 MB）降到 **0**。
-- `GOGC=50`：把 Go 堆增长目标减半。**单独使用无效**（省下的堆会立即被更多 glibc
-  arena 吃掉，实测 20.3 → 21.5 MB 反而变大），必须与 `MALLOC_ARENA_MAX` 配合，
-  两者同时开启才降 **~19%**。网关是 I/O 密集型（本机 9 小时仅消耗 1min10s CPU），
-  多出的 GC 周期与它的空闲 CPU 相比可忽略。
-
-> 桌面版（Electron）已在 `cmd/gui/main.js` 里默认为内嵌核心注入这两个开关；手动 export 同名
-> 环境变量可覆盖。如果你用自己的 systemd unit / 容器，建议照上面加上。
-
-可选：`GOMEMLIMIT=48MiB` 作软上限，实测再省 ~1 MB；代价是逐近上限时 GC 转激进，
-温和场景不必开。
-
-### disable_thinking
-
-请求体带 `"disable_thinking": true` 时，网关透传给各适配器；DeepSeek 适配器将其
-映射为 `extra_body.thinking.type = "disabled"` 关闭推理，其余源按各自协议处理。
-
-### WebUI
-
-内置管理界面（`GET /`），登录后可在浏览器完成：
-
-- **状态页**：源在线状态（实时探测 + 悬停看错误）、模型/源/key 用量统计、请求记录，
-  支持按时间范围导出 CSV；点击模型可生成 pin 到该模型的连接配置。
-- **对话页**：流式/非流式调试。
-- **密钥页**：创建/编辑网关 key，为每个 key 配模型范围（模型 + 源 + token 配额 + 周期），
-  管理员管理全部 key，用户只看到自己的 key。每个模型砖块显示自己的配额徽标
-  （如 `1.0K 77×·1h`，未设配额显示 `∞`），点开可编辑该模型的 token 配额、
-  请求数配额与重置周期 —— 配额按模型独立生效，一个用满不影响同一 key 的其它模型。
-  「我的密钥」页对用户展示本 key 的模型范围与各自配额。
-- **优先级页**：拖拽积木配置 AUTO 链档位。
-- **源页**：在线增删改上游源（API key 等敏感字段加密落盘）。
-- **适配器页**：上传 / 删除 Lua 适配器脚本。
-
-改动写入 `runtime_file`（重启仍生效）。
-
-## Lua 适配器协议
-
-完整 API 见 **[Lua 适配器 API 文档](docs/lua-adapters.md)**（[English](docs/lua-adapters-en.md)）。
-
-每个适配器是一个返回 table 的 Lua 脚本（`internal/lua/adapters/<name>.lua`），
-加载时可被 `adapter_dir` 下的同名脚本覆盖——**改 Lua 脚本无需重编译**，
-重启即生效；通过 WebUI 上传的适配器与在线编辑的源配置则即时生效。
-
-```lua
-return {
-  name = "mysrc",
-  version = "1.0.0",
-  endpoint = "/chat/completions",        -- 上游路径（可被 source.endpoint 覆盖）
-  headers = { ["X-Static"] = "v" },      -- 静态默认请求头（build_headers 缺省时使用）
-
-  -- 请求转换：把 OpenAI 格式 req 转成上游原生格式，返回字符串
-  transform_request = function(raw_json) ... end,
-
-  -- 响应转换：把上游原生响应转成统一格式字符串
-  --   { content, reasoning_content, finish_reason, token_usage{...}, tool_calls[{...}] }
-  transform_response = function(raw_json) ... end,
-
-  -- 流式分块转换：把上游 SSE data 转成 { content, done, ... }，返回 "" 则跳过
-  transform_stream_chunk = function(raw_chunk) ... end,
-
-  -- [可选] 动态请求头/签名钩子
-  -- meta = { url, method, body, api_key, timestamp, source={ name, meta={...} } }
-  build_headers = function(meta) return { ["X-App-Sign"] = sign } end,
-}
-```
-
-内置信号辅助：`hmac_sha256_hex(key, data)`、`sha256_hex(data)`、`base64_encode(s)`、
-`tohex(s)`、`json.encode/decode`、`log(level, msg)`。
-
-### 内置适配器
-
-`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`。
-
-`anthropic`/`gemini`/`ollama` 适配器内置多模态转换（`image_url` → 各自上游格式）；若
-源启用 `disable_thinking`，`deepseek` 适配器会把 `extra_body.thinking.type` 置为
-`disabled`。
-
-**kimicode** 是展示 `build_headers` 的样例：云端校验调用方 app，需要按
-`meta.app_secret` 对时间戳+URL+请求体哈希做 HMAC 签名并附 `X-App-Sign` 等头。
-配好 `sources[].meta.{app_id, app_secret, app_agent}` 即可。
-
-## 桌面 GUI（Electron）
-
-可选独立产物，面向内网桌面端用户：**内嵌 ModelRouter 核心**（Clash Verge 形态），
-一键启停、系统托盘常驻、开机自启、静默启动，内容区一比一内嵌完整 WebUI
-（状态/对话/密钥/优先级/源/适配器 六个页面，免登录）。服务器用户继续使用纯 Go 二进制。
-
-### 两种安装包：Headless 与 Desktop
-
-发布到 Release 的安装包分为两类，面向不同用户：
-
-- **Headless（服务器版）**：纯核心二进制（无 Electron），配置驱动，适合服务器、
-  容器、systemd 等无人值守场景。发布包含 `deb` + `rpm`（含 systemd unit、内存调优
-  与示例配置）与 `tar.gz`（二进制 + `config.example.yaml` + `adapters/` + README），
-  Windows 为 `.zip`。
-- **Desktop（桌面版）**：Electron GUI 安装包，内嵌核心，适合个人桌面日常使用。
-
-### 选择 GUI 还是纯后端
-
-| 场景 | 推荐 | 理由 |
-| ------ | ------ | ------ |
-| 服务器 / 内网网关 / 无人值守常驻 | **Headless**（单二进制） | 极轻量（二进制 ~8MB；RSS 随源数量，单源 ~10MB、16 源 ~32–35MB，见[内存占用](#内存占用实测与调优)），零依赖单进程，直接跑在 systemd/任意容器里，远程管理 |
-| 个人桌面日常使用 / 多设备内网共享 | **Desktop**（GUI） | 免登录内嵌 WebUI、系统托盘一键启停、开机自启、静默后台，适合不懂命令行的使用者 |
-| Windows 桌面 | **Desktop** | 纯后端在 Windows 上需自行注册服务，GUI 提供原生托盘/自启体验 |
-| CI 一键出三平台安装包 | **Desktop 打包脚本** | `make gui-*` 系列出 deb/AppImage/NSIS，可进发行流水线 |
-
-两者完全同源：GUI 内嵌的就是纯后端的同一份 `llmsproxy` 核心（LuaJIT 版），
-配置/适配器格式完全一致，可随时互换。
-
-### 构建与运行
-
-```bash
-cd cmd/gui
-npm install
-make gui          # 或 npm run dev —— 开发运行（需先 make build 生成 bin/llmsproxy）
-make gui-deb      # 本机构建 deb（分享给其它 Linux 用户）
-make gui-dist     # 构建 deb + rpm + AppImage（linux，rpm 需系统 rpmbuild）
-make core-dist    # 构建核心 deb + rpm + tar.gz（服务器版，需 nfpm）
-make gui-win      # 构建 Windows nsis 安装包（本机 mingw + wine）
-make gui-win-docker  # 构建 Windows nsis 安装包（全 docker 自足，含 wine）
-make gui-dist-dir # 仅解包产物（调试用，不产安装包）
-```
-
-产物在 `cmd/build/gui-dist/`：`ModelRouter-<ver>.AppImage`、
-`modelrouter-gui_<ver>_amd64.deb`、`ModelRouter Setup <ver>.exe`（win）；
-核心包在 `cmd/build/dist/`：`llmsproxy_<ver>_amd64.deb`、`llmsproxy-<ver>.x86_64.rpm`。
-rpm 需系统 `rpmbuild`。每次打包后都会跑 `packaging/verify-dist.sh` 产物体检
-（大小下限门禁——1.3.0 曾因宿主 wine 损坏产出 264KB 的残缺 exe 却没有被拦住）。
-发布时会把产物重命名为 `ModelRouter-Desktop-*`（GUI）与 `ModelRouter-Headless-*`（纯后端）
-两类上传到 GitCode Release。
-
-#### Windows 打包（docker 固化方案，推荐）
-
-`make gui-win-docker` 一步完成：交叉编译核心 + NSIS 打包**全在 docker 内**，宿主机
-**无需 mingw、无需 wine、无需 node**。镜像内置 wine（含 i386 运行时），修复了
-此前宿主 wine 损坏（缺 syswow64 → `c0000135`）导致产出 264KB 残缺 exe 的问题。
-
-- 构建镜像：`cmd/gui/docker/win-builder/Dockerfile`（`golang:1.25` + mingw-w64 +
-  node + wine32/wine64 + 预编译 LuaJIT windows 版 `lua51.dll` 与 import lib）。
-  LuaJIT 源码取自 gitcode 镜像（国内可达），github 兜底；Go 依赖走 `https://goproxy.cn`。
-- 一键发布：`cmd/gui/scripts/dist-win-docker.sh` —— 核心编译 + electron-builder
-  NSIS 全部在容器内执行（`USE_SYSTEM_WINE=true` 走镜像内的 wine）；electron-builder
-  的 ~400MB 工具链缓存挂到 docker volume，二次构建增量复用。
-- 首次运行自动构建 `modelrouter/win-builder` 镜像（约 5–10 分钟）；已存在则直接复用。
-- 打包完成后自动跑 `packaging/verify-dist.sh` 门禁：exe 低于 5MB / deb·rpm·7z 低于
-  10MB 即失败——残缺安装包不可能再溜进 Release。
-
-```bash
-# 全 Windows 发布（核心重编 + NSIS）：
-make gui-win-docker
-```
-
-### 特性
-
-- **内嵌核心**：自动拉起捆绑的 `llmsproxy`（LuaJIT 版），配置/密钥/adapter 存于
-  用户数据目录 `<userData>/profile/`，首次运行自动生成随机 admin key 写入 `keys`
-  （非 seed）并自注入，WebUI 免登录、不会提示"更换初始密钥"。端口可在设置中更改（默认 8787）。
-- **系统托盘**：内核状态点、开机自启/静默启动开关、启停/重启内核、退出。
-- **静默启动**：`--silent` 参数或设置项，启动仅驻托盘不弹窗；首次运行始终显示窗口。
-- **开机自启**：Windows/macOS 用 `setLoginItemSettings`；Linux 写
-  `~/.config/autostart/modelrouter-gui.desktop`，静默开启时自动附加 `--silent`。
-
-## 目录
-
-```
-cmd/llmsproxy            # 入口（纯 Go 服务端）
-cmd/gui                  # Electron 桌面 GUI（内嵌核心 + 托盘 + 打包）
-internal/config          # YAML 配置加载/校验
-internal/lua             # LuaJIT VM + worker 池 + AdapterCache + 内置适配器 (embed)
 internal/provider        # Provider(HTTP) + Registry(路由)
-internal/gateway         # OpenAI 兼容 HTTP 服务 + 鉴权 + SDK/流式 + WebUI
+internal/gateway         # OpenAI 兼容 HTTP 服务 + 鉴权 + 流式 + WebUI
 internal/types           # 统一格式 & OpenAI wire 类型
+packaging/               # deb/rpm/tar 打包、systemd unit、迁移脚本
+docs/                    # 主题文档
 ```
 
-## 开发协作
+## 开发与测试
-
-- **Git 分支工作流**：主分支 / 特性分支 / 发布分支的规范见
-  **[docs/git-workflow.md](docs/git-workflow.md)**（[English](docs/git-workflow-en.md)）。
-  主线：`main` 永可部署 → 特性合入 `main` → 切 `release/vX.Y.Z` 打 tag 发布 →
-  hotfix 提交发布分支并 cherry-pick 回 `main`。
-- **Lua 适配器协议**：见 **[docs/lua-adapters.md](docs/lua-adapters.md)**
-  （[English](docs/lua-adapters-en.md)）。
-
-## 测试
 
 ```bash
-go test -tags luajit ./...
+go test -tags luajit ./...          # 全量
+go test -tags luajit ./internal/lua # 单包（LuaJIT 相关改动必跑）
 ```
 
-覆盖：配置校验、适配器加载/变换、签名钩子、Gateway 鉴权、SDK 结算、SSE 流式、
+- 分支与发布纪律见 [docs/git-workflow.md](docs/git-workflow.md)。
-模型路由、多模态透传与 disable_thinking。
+- 插件放置在 `plugin_dir`，网关启动时**不会**创建、覆盖或删除任何插件文件；
+  升级桌面端也不会动它们。
+
+## 许可
+
+MIT License，见 [LICENSE](./LICENSE)。

README_EN.md

@@ -1,626 +1,155 @@
 # ModelRouter
 
+> **Repository**: <https://atomgit.com/HeNanMU-EEA/ModelRouter>
+>
 > **中文**: [README.md](./README.md)
 >
-> **AI-assisted**: This project is developed with AI-assisted coding (code and docs are produced in collaboration with AI, reviewed by humans).
+> **AI-assisted**: code and documentation were produced with AI assistance and
+> reviewed by a human.
 
-A lightweight, unified OpenAI-compatible LLM gateway for internal networks —
+An intranet gateway for people with several LLM subscriptions: configure once and
-configure once, and let multiple services (agents, SDKs, bots) share many
+let multiple services share the upstream capacity. Pick a model explicitly, or let
-upstreams (DeepSeek, Qijiar, OpenAI, Anthropic, Gemini, Groq, Mistral, Ollama,
+`AUTO` choose by priority among the sources that are currently healthy.
-KimiCode…) behind a single endpoint. Pick a specific model or use **AUTO** mode,
-which routes to the best healthy upstream by configured priority.
 
-> **Lightweight, no recompile to add sources**: the gateway is a single Go
+The gateway itself is a single Go binary (~10 MB, no runtime dependencies). Adding
-> binary (8-12 MB, ~8 MB stripped; zero runtime dependencies). Adding or switching an upstream is
+or switching an upstream is a `sources` entry in `config.yaml` (or the WebUI), or a
-> just a `sources` entry in `config.yaml` (or via the WebUI) or a `.lua`
+dropped-in `.lua` adapter — **no Go changes, no rebuild**. Changes submitted through
-> adapter — **no Go changes, no recompile**. Changes submitted through the
+the WebUI take effect immediately; editing `config.yaml` or a `.lua` adapter file
-> WebUI take effect immediately (hot reload); editing `config.yaml` or a `.lua`
+directly requires a restart.
-> file under `adapter_dir` by hand requires a process restart.
 
-## Overview
+## What it does
 
-Exposes a standard `OpenAI Chat Completions` API (`/v1/chat/completions` +
+- **One OpenAI-compatible API**: `/v1/chat/completions`, `/v1/models`,
-`/v1/models`) to your internal network, translating between multiple upstream
+  `/v1/images/generations`. Every source speaks the same shape outwards, including
-protocols via **Lua adapters**. Supports one-shot calls, SSE streaming,
+  `reasoning_content`, `tool_calls` and `usage`.
-**AUTO model routing**, multimodal passthrough, and tool calls.
+- **Many sources, one AUTO chain**: configure any number of upstreams; `AUTO`
-
+  follows the priority tiers saved in the WebUI, fails over between them, cools
-Extracted and independently evolved from the multi-source LLM adapter layer of
+  down broken sources and probes them back to health at half-cool.
-[HomeAgent](https://gitcode.com/JianFeeeee/HomeAgent)
+- **Protocol differences live in Lua**: each source binds a `.lua` adapter that
-(`internal/agent/api/provider.go` + `internal/lua/adapters/*`).
+  transforms requests, responses and stream chunks both ways, and can sign headers
-
+  via `build_headers(meta)`. Vendor knowledge stays out of the Go kernel.
-## Features
+- **Plugins are independent**: business extensions (channel login, billing, quota
-
+  probing) are external Lua plugins loaded from `plugin_dir`. They are not compiled
-- **Multi-source**: any number of upstream sources in one process, routed by the
+  into the binary and not shipped in release archives, and they talk to the kernel
-  request's `model`.
+  only through the `modelrouter` SDK.
-- **AUTO mode**: with `default_model: AUTO`, scheduling follows the tiered AUTO
+- **Multi-tenant keys**: gateway keys are separate from upstream keys, and can be
-  chain saved on the Priority page (see below).
+  scoped to a model set, an AUTO chain and a quota.
-- **Unified output**: every source speaks OpenAI format (incl.
+- **Encrypted at rest**: upstream keys, custom header values and gateway keys are
-  `reasoning_content`, `tool_calls`, `usage`).
+  sealed with AES-256-GCM; `master.key` is 0600, or injected via
-- **Image generation**: `POST /v1/images/generations`, routed to models with
+  `MODELROUTER_MASTER_KEY`.
-  `kind: image`.
+- **Observability**: live source probing, HTTP access log, config-change audit,
-- **Per-model quota**: each key gives every model its own token and request caps
+  request and usage CSV export. Writes are kept off the request hot path.
-  plus a reset period (hour/week/month/custom N hours). One model running out
-  affects only that model — the key's other models keep working. Exhaustion
-  answers 429 + `Retry-After` naming the model, so a client resumes when the
-  window rolls over; admin keys are never capped.
-- **Multimodal**: `content` arrays (`image_url` etc.) pass through losslessly;
-  Anthropic/Gemini/Ollama are translated automatically.
-- **LuaJIT VM**: golua-binding LuaJIT; each adapter has its own VM + worker
-  pool for safe concurrency. Pools are **elastic**: the sum of `max_concurrent`
-  over an adapter's sources is a ceiling, not a preallocation — states are booted
-  on demand and reclaimed when demand drops, so an idle gateway holds close to
-  zero Lua states. The grow step follows the adapter's max concurrency
-  (`clamp(ceil(max/8), 1, 8)`, and never exceeds the number of queued callers);
-  the shrink step follows the live connection count
-  (`clamp(ceil(slack/(1+in_use)), 1, slack)`), so an idle pool collapses in one
-  round while a busy one gives up a single state at a time.
-- **On-demand request logs**: dashboard totals are computed from the **full**
-  audit history (streamed once at startup and released — 29 MB / 221k lines in
-  ~260 ms), while the raw records are never held in memory: the dashboard loads
-  one screen, scrolling pages the rest straight off disk, CSV export streams in
-  O(1) memory, and leaving the page releases everything.
-- **Self-healing cooldown**: cooldown is capped at 5 minutes and, past the
-  window's midpoint, exactly one probe request is allowed through; a recovered
-  upstream (or a reset quota) returns to full rotation on that probe instead of
-  waiting out the window.
-- **disable_thinking**: `disable_thinking: true` toggles reasoning per-request.
-- **Lua adapter protocol**: each source mounts a `.lua` adapter with
-  `transform_request` / `transform_response` / `transform_stream_chunk` — all
-  protocol differences live in Lua, Go only schedules and proxies.
-- **Signature / header hooks**: adapters may define `build_headers(meta)` to
-  inject/sign request headers before the HTTP call (e.g. KimiCode style
-  app-validation), with helpers like `hmac_sha256_hex`, `sha256_hex`,
-  `base64_encode`.
-- **WebUI**: built-in management page to view/add/edit sources & models, edit
-  the AUTO chain, and manage per-key model scopes, persisted to a runtime file.
-- **Encrypted secrets at rest**: upstream `api_key`, custom header values and
-  gateway keys are stored AES-256-GCM encrypted (`master.key` 0600 next to the
-  runtime file, or `LLMS_PROXY_MASTER_KEY`).
-- **Live source probing**: the status page probes each source
-  (`GET {base}/models`, falling back to a minimal chat call) without touching
-  the scheduler's backoff state; errors are shown on hover.
-- **Auth**: the gateway validates client keys via `gateway_keys`; independent
-  from each upstream's own key.
-- **Streaming**: SSE `chat.completion.chunk` with a role-first chunk and
-  `[DONE]` terminator.
 
 ## Quick start
 
 ```bash
-GOMODCACHE=... GOPROXY=off go build -tags luajit -o llmsproxy ./cmd/llmsproxy
+# LuaJIT development libraries are required; -tags luajit is mandatory
-./llmsproxy -config config.yaml    # first run generates a default config and prints a random admin key
+GOMODCACHE=... GOPROXY=off go build -tags luajit -o modelrouter ./cmd/modelrouter
+
+# First run generates a default config and prints a random admin key
+./modelrouter -config config.yaml
 ```
 
-> Depends on [golua](https://github.com/aarzilli/golua) (LuaJIT bindings).
+> Depends on [golua](https://github.com/aarzilli/golua) (LuaJIT bindings). Without
-> You **must** build with `-tags luajit`; otherwise the built-in superset
+> `-tags luajit` the build falls back to gopher-lua, which behaves slightly
-> gopher-lua path is used (behavior differs slightly).
+> differently.
 
-> **No config is shipped in the repo** (config files carry real keys). On
+> **No config in the repo**: the repository ships no config file (they contain
-> first run the binary generates a default config at the `-config` path: a
+> keys). The generated default binds `127.0.0.1:8080` only; rotate the admin key in
-> random admin key (printed to the startup log), loopback-only `127.0.0.1:8080`,
+> the WebUI after the first login.
-> and a keyless source ready to chat. Rotate the admin key in the
-> WebUI after first login.
 
 ```bash
-# no key -> 401
+# No key -> 401
 curl http://127.0.0.1:8080/v1/models
 
-# one-shot ($KEY = the admin key printed at first startup)
+# Non-streaming ($KEY = the admin key printed at first start)
 curl -H "Authorization: Bearer $KEY" \
      -d '{"model":"AUTO","messages":[{"role":"user","content":"hi"}]}' \
      http://127.0.0.1:8080/v1/chat/completions
 
-# streaming
+# Streaming
 curl -N -H "Authorization: Bearer $KEY" \
      -d '{"model":"AUTO","stream":true,"messages":[{"role":"user","content":"hi"}]}' \
      http://127.0.0.1:8080/v1/chat/completions
 ```
 
-Any OpenAI SDK works: point `base_url` at the gateway, use any authorized
+Any OpenAI SDK works: point `base_url` at the gateway and use an authorized gateway
-gateway key as the API key.
+key as `api_key`.
-
+
-## Configuration
+## Documentation
-
+
-Generated on first run at the `-config` path (default `./config.yaml`). Core fields:
+The detailed documentation is currently written in Chinese.
-
+
-```yaml
+| Topic | Document |
-listen: 127.0.0.1:8080        # bind address (keep internal/loopback)
+|---|---|
-gateway_keys: [sk-gw-<random>] # initial admin key seed (random per install), migrated to the store on first start
+| Implementation and the full feature list | [docs/features.md](docs/features.md) |
-default_model: AUTO           # when model is unroutable, follow the AUTO chain
+| Configuration: `config.yaml`, model discovery, multiple keys | [docs/configuration.md](docs/configuration.md) |
-adapter_dir: adapters         # Lua adapter dir; created+seeded if missing, read-only otherwise
+| Routing: model routing, AUTO chain, cooldown, sources vs adapters, source templates | [docs/routing.md](docs/routing.md) |
-runtime_file: runtime.json    # WebUI-edited sources/keys/AUTO chain persist here
+| Adapters: Lua protocol and built-in adapters | [docs/adapters.md](docs/adapters.md) · [docs/lua-adapters-en.md](docs/lua-adapters-en.md) (English) |
-
+| Plugins: SDK contract, hooks, UI extensions, HTTP endpoints | [docs/plugins.md](docs/plugins.md) |
-sources:
+| Desktop app: Headless / Desktop artifacts, build and run | [docs/gui.md](docs/gui.md) |
-  - name: deepseek
+| Performance: measured memory use, tuning, thinking budget | [docs/performance.md](docs/performance.md) |
-    base_url: https://api.deepseek.com
+| WebUI usage | [docs/webui.md](docs/webui.md) |
-    api_key: sk-...            # or api_key_env: SOME_ENV to read from an env var
+| Development: layout, tests, branch and release discipline | [docs/development.md](docs/development.md) · [docs/git-workflow-en.md](docs/git-workflow-en.md) (English) |
-    adapter: deepseek
+| Deployment and migration | [docs/deployment.md](docs/deployment.md) |
-    max_concurrent: 8
+
-    models:
+## Renamed from llmsproxy
-      - id: deepseek-v4-flash
+
-        priority: 100          # YAML sources seed the AUTO chain on first start
+The product name has been ModelRouter for a while; the binary, config directory,
-        kind: chat
+service name and environment variables are now `modelrouter` as well:
-      - id: deepseek-v4-pro
+
-        priority: 60
+| Before | After |
-        kind: chat
+|---|---|
-    # static headers (take precedence over adapter defaults)
+| `llmsproxy` (binary) | `modelrouter` |
-    headers: { X-Tenant: prod }
+| `/etc/llmsproxy/` | `/etc/modelrouter/` |
-    # passthrough metadata for the Lua build_headers hook
+| `llmsproxy.service` | `modelrouter.service` |
-    meta: { app_id: x, app_secret: y }
+| `LLMSPROXY_*` | `MODELROUTER_*` |
-    temperature: 0.7
+
-    max_tokens: 4096
+Existing installs migrate with a dry-run-by-default script (copies, never moves,
-    timeout: 120s              # request timeout, default 120s
+and can roll back):
-```
-
-Sensitive fields in the runtime file (`runtime_file`) are encrypted at rest:
-
-- AES-256-GCM, stored as `enc:v1:<base64>`.
-- Master key source: env var `LLMS_PROXY_MASTER_KEY` (64 hex chars), else a
-  `master.key` file next to the runtime file; if neither exists it is generated
-  on first start (mode 0600).
-- Old plaintext files still load; the first UI save migrates the whole file to
-  ciphertext.
-- Back up `master.key` with your config — losing it makes the secrets
-  undecryptable. Do not commit it.
-
-### Gateway keys (multi-key)
-
-Gateway auth uses a "multi-key + role + model scope" model. Keys are persisted
-under the `keys` field of the runtime file (encrypted at rest):
-
-- The `gateway_keys` config is only an **initial admin key seed** — it is
-  migrated into the runtime store on first start and no longer drives auth.
-- **Important: after first start, replace the admin key via the WebUI Keys
-  page.** The seed key is written in plaintext in `config.yaml`, so keeping it
-  active is a security risk; create a new admin key, log in with it, then
-  delete the seed key.
-- The WebUI **Keys page** creates/deletes keys. Each key has a role (`admin`
-  manages everything, `user` sees only its own key) and an optional **model
-  scope** (model + source + token quota + reset period). Each model brick shows
-  its own budget badge (e.g. `1.0K 77×·1h`, `∞` when uncapped); clicking it
-  edits that model's token quota, request quota and reset period. Quotas apply
-  per model, so one model running out never blocks the key's others. The
-  "My key" view shows a user its model scopes and their budgets.
-- Clients authenticate with any authorized key's plaintext as
-  `Authorization: Bearer <key>`.
-- Deleting a key removes it from the store immediately.
-
-#### Per-key usage quota
-
-**Quotas are per model.** Each key's `models[]` list gives every model its own
-token budget and request budget. One model running out affects only that model
-— the key's other models keep working.
-
-```yaml
-keys:
-  - key: sk-gw-<hex>
-    role: user
-    name: agent-alice
-    models:
-      - model: deepseek-v4-flash
-        token_quota: 1000000   # this key's token budget for this model
-        req_quota: 20000       # requests within the window
-        period: nhour          # "" | hour | week | month | nhour
-        hours: 6               # n-hour only
-      - model: AUTO
-        token_quota: 5000000   # AUTO is a quota entry like any other
-        period: hour
-      - model: kimi-k3        # no quota listed = unlimited
-```
-
-- **There is deliberately no key-wide total.** A key-wide cap would make
-  "move A's budget to B" a global reallocation; per-model budgets keep each
-  model independently controllable, so it stays visible which model is
-  actually consuming the spend.
-- `period`: empty = never resets (lifetime total); `hour` / `week` / `month` =
-  fixed windows; `nhour` + `hours` = a custom hour count. **A misspelled
-  period is rejected at write time** rather than silently becoming a
-  never-resetting quota.
-- Quotas are isolated per key *and* per model within a key: one key exhausting
-  `m1` never draws on another key's budget, and never blocks the same key's
-  `m2`.
-- Usage counts chat, streaming and image requests, and survives a restart by
-  replaying the audit log (40 days retained, covering the longest `month`
-  window).
-- An exhausted quota returns **429 + `Retry-After`**
-  (`rate_limit_exceeded`) and the message names the model that ran out, so a
-  client resumes when the window rolls over; a model the key may not use stays
-  403. **Admin keys are never capped** (quotas on their scopes are not
-  enforced either), so a cap can never lock the operator out.
-- Buckets are whole unix hours, so a window frees up at most an hour late
-  (deliberately freeing late rather than overspending).
-- `PUT /api/keys/{key}` submits quotas by submitting `models` — the caps are
-  part of the scope, so there is no second budget that can drift out of sync
-  with the model list. An explicit `0` lifts that model's cap.
-
-##### Quota rejection vs capacity rejection
-
-A quota is checked at the entry point and occupies no upstream slot, so it is a
-cheap rejection. Running out of capacity instead waits out the bounded
-`busyWait` before answering 503. Measured (capacity 8, 0.6 s/request, 100
-concurrent clients):
-
-| Case | Status | Latency | Reaches upstream |
-|---|---|---|---|
-| Quota spent (`req_quota` 50, 100 concurrent) | `429 rate_limit_exceeded` + `Retry-After` | **19–21 ms** | no |
-| Capacity exhausted (same topology, no quota) | `503 upstream_error` | ~2.6 s (`busyWait`) | overflow: no |
-
-So a client can tell "wait for the window to roll over" apart from "wait for
-upstream capacity" — the first only needs patience, the second usually means
-back off concurrency or switch sources.
-
-##### What the quota costs
-
-Buckets are kept per (key, model, whole unix hour) for 40 days. Measured on an
-AMD 7840HS:
-
-- A window query scans the window, not the whole retention: 49 ns for 1 h,
-  55 ns for 24 h, 3.9 µs for 30 d.
-- Recording one request: 283 ns, 3 allocations (unchanged from before this
-  feature; the allocations come from the record ring buffer).
-- Memory: the production shape (7 keys x 8 models x 2 sources at full 40-day
-  retention) costs about **3.7 MB**. The `source::model` bucket used by a
-  source-pinned quota is created **lazily** — it is maintained only once some
-  quota actually pins that source, because writing it on every record would
-  cost an extra 18 MB at 20 keys x 8 models x 3 sources.
-  The trade-off: usage recorded before a pinned quota was configured cannot be
-  split by source afterwards, so **a pinned quota may under-count its first
-  window**.
-
-```
-HTTP/1.1 429 Too Many Requests
-Retry-After: 2100
-{"error":{"type":"rate_limit_exceeded","message":"token quota exceeded for \"deepseek-v4-flash\" (5000000/5000000)"}}
-```
-
-### Model routing
-
-`/v1/chat/completions` `model` resolution order:
-
-1. `source/model` or `source:model` prefix → pinned source;
-2. exact match of a source's `model`;
-3. with `default_model: AUTO` → follow the AUTO chain from the Priority page
-   (see below);
-4. otherwise the request errors.
-
-Any OpenAI client can pin to a source by setting `model` to its
-`name/anything`; `AUTO` (or the gateway's `default_model: AUTO`) follows the
-tiered AUTO chain.
-
-### AUTO chain (Priority page)
-
-AUTO scheduling is driven **only** by the rules saved on the Priority page. The
-numeric `models[].priority` in source configs no longer participates in
-scheduling and is no longer shown. **Chat and image generation are two
-independent chains**, switched in the Priority page via the `Chat / Image` toggle:
-
-- **Chat chain** (`auto` field): serves `POST /v1/chat/completions` with
-  `model: AUTO`; only `kind: chat` models are honored (image slots are skipped).
-- **Image chain** (`auto_image` field): serves `POST /v1/images/generations` with
-  `model: AUTO`; only `kind: image` models are honored. With no image chain
-  configured, AUTO image generation falls back to legacy discovery (all sources
-  exposing an image model, in registry order) for backward compatibility.
-
-Each rule is one "slot": `{ model, source, tier, token_quota, period, hours }`.
-
-- `tier` is the priority tier: models in the same tier sit side by side and
-  share it; tiers run high → low.
-- The same model may appear in several slots (e.g. A low → B low → A high →
-  B high) and is tried in tier order.
-- `token_quota` > 0 means the slot is skipped once its tokens within the reset
-  window are exhausted; `period` supports `hour` / `week` / `month` / `nhour`
-  (with `hours`); empty = unlimited.
-- Within a tier, slots run by preference score; cooling / quota-exhausted /
-  hard-failed slots fall through to the next tier, and total failure returns
-  503 with a per-tier summary.
-
-### Sources vs. adapters
-
-- A **source** describes a connection to an upstream: `name`, `base_url`,
-  `api_key`, model list and priorities.
-- An **adapter** is the protocol translation logic (Lua script): converts the
-  unified OpenAI request to the upstream's native format and back.
-- One adapter serves many sources (e.g. `openai.lua` for any OpenAI-compatible
-  site); one source can switch adapters via its `adapter` field.
-- Sources decide *whom to talk to and which models to expose*; adapters decide
-  *how to talk*. They are linked by the `adapter` field inside each `sources[]`
-  entry.
-
-Loading flow (assembled by `Core` in `internal/core`):
-
-1. On startup `lua.NewVM(adapter_dir)` loads all adapters: built-in ones
-   (embedded at compile time) + same-name override files under `adapter_dir`.
-2. `config.Load` reads `config.yaml`; `config.NewStore(runtime_file)` reads
-   WebUI-edited runtime sources; both are merged by name.
-3. `rebuildRegistry` creates a `provider.Provider` per source (holding its
-   adapter), sizing the adapter's worker pool from the source's concurrency
-   limit.
-4. Incoming requests are routed by `Registry` to a provider, which calls the
-   adapter's `transform_request` → HTTP call → `transform_response` /
-   `transform_stream_chunk`.
-
-"Add/Edit source", "Upload adapter", "AUTO chain edits" and "per-key model
-scopes" in the WebUI take effect immediately (written to the runtime file /
-`adapter_dir`, then reassembled — no restart). Directly editing `config.yaml`
-or files under `adapter_dir` requires a process restart.
-
-Deleting a source/adapter in the WebUI truly removes runtime (UI-created)
-sources and `.lua` files; base sources defined in `config.yaml` cannot rewrite
-that file, so they are hidden via a deletion tombstone (still hidden after
-restart) and can be restored by re-adding the same name in the UI.
-
-### Source templates (multi-key balancing)
-
-When several upstream keys share one config (URL / adapter / model list /
-concurrency / rpm), copying the whole source block N times is wasteful. A
-**source template** stores every source field except `name` and `api_key`, so you
-create N key-only-different sources from one shared skeleton — natural multi-key
-load balancing (each expanded source schedules, cools down, and reports health
-independently).
-
-- WebUI Sources page, top-right **Templates** button: list all templates with
-  edit / delete / create.
-- Add-source dialog has two header buttons:
-  - **From template** → pick a template → form auto-fills (name + key still
-    yours to enter);
-  - **As template** → save the current form's non-key/name fields as a template.
-- Templates persist in the runtime file's `source_templates` field alongside
-  runtime sources (hot-managed, apply immediately). Templates hold no secret
-  keys, so they're safe to share / version-control.
-
-A template is just a recipe — it doesn't become a source by itself. Only the
-key-bearing sources created from it carry real traffic.
-
-### Memory footprint (measured and tuned)
-
-Memory is not a constant: it scales with the **number of configured sources**
-(each source owns an `http.Transport` with its connection pool plus per-model
-health state), not with uptime. Measured on this machine (Linux x86_64, 12 cores):
-
-| Deployment shape | Startup RSS | Settled RSS |
-|---|---|---|
-| 1 source / 1 adapter (minimal) | ~4 MB | ~10 MB |
-| 1 source + a 29 MB audit history | ~19 MB | ~20 MB |
-| **16 sources / 13 adapters / 59 models (this host)** | ~28 MB | **~32-35 MB** |
-
-> For reference, the same production config used **~105 MB** before this round of
-> work. The reduction comes from three places: raw audit records are no longer kept
-> resident (~25 MB — the aggregates still scan the full history, but the scan
-> releases as it goes), Lua state pools no longer grow monotonically, and the two
-> runtime knobs below.
-
-Breakdown of the production instance (per-region, from `/proc/<pid>/smaps`):
-
-| Region | RSS | Notes |
-|---|---|---|
-| Go heap | ~14 MB | provider/registry/scheduler state + connection-pool buffers |
-| Other anonymous (thread stacks / LuaJIT chunks / runtime) | ~12 MB | scales with thread count and loaded adapters |
-| Binary text+rodata | ~8 MB | mapped executable pages (read-only, reclaimable by the kernel) |
-| Go runtime reservations | ~4 MB | the 2 GB+ you see in VSZ is address space, not physical memory |
-| Shared libraries | ~3 MB | libc / libluajit / libm |
-
-**Two recommended deployment knobs** (environment only, no code change):
-
-```ini
-# /etc/systemd/system/llmsproxy.service
-Environment=GOGC=50
-Environment=MALLOC_ARENA_MAX=2
-```
-
-- `MALLOC_ARENA_MAX=2`: LuaJIT allocates through cgo into glibc malloc, and glibc
-  allows up to `8 x nproc` per-thread arenas. Every OS thread that touches malloc
-  claims one (~1 MB each) and **never returns it to the OS**. Measured: 8-15
-  arenas (7-12 MB) down to **0**.
-- `GOGC=50`: halves the Go heap growth target. **It does nothing on its own** —
-  the heap it saves is immediately consumed by additional glibc arenas (measured
-  20.3 -> 21.5 MB, i.e. slightly worse) — so it must be paired with
-  `MALLOC_ARENA_MAX`. Together they cut settled RSS by **~19%**. The gateway is
-  I/O bound (1min10s of CPU per 9 hours here), so the extra GC cycles are free.
-
-> The desktop build (Electron) already injects both when spawning the embedded
-> core (`cmd/gui/main.js`); exporting either variable yourself overrides it. If you
-> write your own systemd unit or container spec, add them there.
-
-Optional: `GOMEMLIMIT=48MiB` as a soft ceiling saves roughly 1 MB more, at the
-cost of GC turning aggressive as the limit approaches. Not worth it for a relaxed
-deployment.
-
-### disable_thinking
-
-With `"disable_thinking": true` in the request body, the gateway passes it to
-each adapter; the DeepSeek adapter maps it to `extra_body.thinking.type =
-"disabled"`, other sources follow their own protocol.
-
-### WebUI
-
-Built-in admin page at `GET /`; after login you can:
-
-- **Status**: source online state (live probe, error on hover), per
-  model/source/key usage stats, request records filtered by key, and CSV export
-  over a time range; click a model to pin the connection snippet to it.
-- **Chat**: streaming / non-streaming debug.
-- **Keys**: create/edit gateway keys, set per-key model scopes (model + source +
-  token quota + reset period); admins manage all keys, users see only their own.
-- **Priority**: drag blocks to build the tiered AUTO chain.
-- **Sources**: add/edit/delete upstream sources online (secrets stored
-  encrypted).
-- **Adapters**: upload / delete Lua adapter scripts.
-
-Changes persist to `runtime_file` (survive restarts).
-
-## Lua adapter protocol
-
-Full API: **[Lua adapter docs](docs/lua-adapters.md)**
-([中文](docs/lua-adapters-en.md)).
-
-Each adapter is a Lua script returning a table
-(`internal/lua/adapters/<name>.lua`), overridable by a same-name file under
-`adapter_dir` — **no recompile needed**, restart to take effect; uploads via
-the WebUI take effect immediately.
-
-```lua
-return {
-  name = "mysrc",
-  version = "1.0.0",
-  endpoint = "/chat/completions",        -- upstream path (source.endpoint overrides)
-  headers = { ["X-Static"] = "v" },      -- static default headers (fallback)
-
-  -- request: convert unified OpenAI req -> upstream native format, return string
-  transform_request = function(raw_json) ... end,
-
-  -- response: convert upstream response to unified format string
-  --   { content, reasoning_content, finish_reason, token_usage{...}, tool_calls[{...}] }
-  transform_response = function(raw_json) ... end,
-
-  -- streaming chunk: convert upstream SSE data to { content, done, ... }; "" skips
-  transform_stream_chunk = function(raw_chunk) ... end,
-
-  -- [optional] dynamic headers / signature hook
-  -- meta = { url, method, body, api_key, timestamp, source={ name, meta={...} } }
-  build_headers = function(meta) return { ["X-App-Sign"] = sign } end,
-}
-```
-
-Shared helpers: `hmac_sha256_hex(key, data)`, `sha256_hex(data)`,
-`base64_encode(s)`, `tohex(s)`, `json.encode/decode`, `log(level, msg)`.
-
-### Built-in adapters
-
-`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama`
-`kimicode`.
-
-`anthropic`/`gemini`/`ollama` include multimodal conversion
-(`image_url` → their native format); with `disable_thinking` the `deepseek`
-adapter sets `extra_body.thinking.type` to `disabled`.
-
-**kimicode** demonstrates `build_headers`: the cloud validates the calling
-app, so you HMAC-sign timestamp+URL+body with `meta.app_secret` and add
-`X-App-Sign`-style headers. Configure `sources[].meta.{app_id, app_secret,
-app_agent}`.
-
-## Desktop GUI (Electron)
-
-Optional standalone product for desktop users on your internal network: an
-**embedded ModelRouter core** (Clash Verge-style): one-click start/stop, system
-tray, autostart, silent launch, and a pixel-for-pixel embedded full WebUI
-(Status/Chat/Keys/Priority/Sources/Adapters — no login needed). Server users
-keep using the plain Go binary.
-
-### Two installer kinds: Headless and Desktop
-
-Release installers ship in two flavors for different audiences:
-
-- **Headless (server)**: the plain core binary (no Electron), config-driven,
-  for servers / containers / systemd / unattended runs. Ships as `deb` + `rpm`
-  (systemd unit with memory tuning + example config) and `tar.gz` (binary +
-  `config.example.yaml` + `adapters/` + README), Windows as `.zip`.
-- **Desktop (desktop app)**: the Electron GUI installer with an embedded core,
-  for personal daily desktop use.
-
-### GUI vs. plain backend
-
-| Scenario | Pick | Why |
-| ---- | ---- | ---- |
-| Server / intranet gateway / unattended long-running | **Headless** (single binary) | ~8 MB binary; RSS scales with the number of sources (~10 MB for one, ~32-35 MB for 16 — see [Memory footprint](#memory-footprint-measured-and-tuned)), zero-dep single process — drop it into systemd or any container, remote admin |
-| Personal desktop daily use / multi-device intranet sharing | **Desktop** (GUI) | no-login embedded WebUI, tray one-click, autostart, silent background — for non-CLI users |
-| Windows desktop | **Desktop** | plain backend needs manual service registration; GUI ships native tray/autostart |
-| CI one-shot 3-platform installers | **Desktop packaging scripts** | `make gui-*` emits deb / AppImage / NSIS, drops straight into a release pipeline |
-
-Both share the exact same `llmsproxy` core (LuaJIT build) — configs and
-adapters are fully compatible, freely interchangeable.
-
-### Build & run
 
 ```bash
-cd cmd/gui
+sudo ./packaging/migrate-to-modelrouter.sh              # preview
-npm install
+sudo ./packaging/migrate-to-modelrouter.sh --apply      # apply
-make gui            # or npm run dev — dev run (needs make build first for bin/llmsproxy)
+sudo ./packaging/migrate-to-modelrouter.sh --rollback --apply
-make gui-deb        # local deb (share with other Linux users)
-make gui-dist       # deb + rpm + AppImage (linux; rpm needs system rpmbuild)
-make core-dist      # core deb + rpm + tar.gz (server edition; needs nfpm)
-make gui-win        # Windows NSIS (host mingw + wine)
-make gui-win-docker # Windows NSIS, fully dockerized (wine in image; no host mingw/node/wine)
 ```
 
-Artifacts land in `cmd/build/gui-dist/`: `ModelRouter-<ver>.AppImage`,
+Old paths and old environment variable names are still accepted on the read side,
-`modelrouter-gui_<ver>_amd64.deb`, `ModelRouter Setup <ver>.exe` (win); core
+so an upgrade cannot fail merely because of ordering. `deploy.sh` refuses to run
-packages land in `cmd/build/dist/`: `llmsproxy_<ver>_amd64.deb`,
+when it finds the old `/etc/llmsproxy` without a new config, because that would
-`llmsproxy-<ver>.x86_64.rpm`. Both target families run
+deploy a gateway with no key and no sources.
-`packaging/verify-dist.sh` afterwards — a size floor that rejects degenerate
-installers (1.3.0 shipped a 264 KB NSIS stub because the host wine was broken
-and nothing caught it). At release time artifacts are renamed into the
-`ModelRouter-Desktop-*` (GUI) and `ModelRouter-Headless-*` (plain backend)
-families before upload to GitCode Releases.
 
-#### Windows packaging (dockerized, recommended)
+The plugin SDK follows the same naming: the only public entry point is
-
+`modelrouter` (`llmsproxy` remains a compatibility alias). See
-`make gui-win-docker` does cross-compile + NSIS **entirely in docker**; the host
+[docs/plugins.md](docs/plugins.md).
-needs **no mingw, no wine, no node**. Wine (with the i386 runtime) lives inside
- the image, fixing the broken-host-wine failure that produced the 264 KB stub
-installer.
-
-- Image: `cmd/gui/docker/win-builder/Dockerfile` (`golang:1.25` + mingw-w64 +
-  node + wine32/wine64 + prebuilt LuaJIT for Windows: `lua51.dll` + import lib).
-  LuaJIT source comes from the gitcode mirror (reachable in CN) with github
-  fallback; Go modules go through `https://goproxy.cn`.
-- One-shot release: `cmd/gui/scripts/dist-win-docker.sh` — core build + NSIS
-  packaging run inside the container (`USE_SYSTEM_WINE=true` uses the image's
-  wine); electron-builder's ~400 MB toolchain cache lives in a docker volume so
-  later builds reuse it.
-- The `modelrouter/win-builder` image is auto-built on first run (~5-10 min)
-  and reused afterwards.
-- After packaging, `packaging/verify-dist.sh` runs as a gate: an exe under
-  5 MB (or a deb/rpm/7z under 10 MB) fails the build — a broken installer can
-  never reach a Release again.
-
-```bash
-# Full Windows release (rebuild core + NSIS):
-make gui-win-docker
-```
-
-### Features
-
-- **Embedded core**: auto-spawns the bundled `llmsproxy` (LuaJIT build);
-  config/keys/adapters live in `<userData>/profile/`. First run generates a
-  random admin key written into `keys` (not seed) and self-injects it — WebUI
-  needs no login and never nags about changing the initial key. Port changeable
-  in settings (default 8787).
-- **System tray**: core status dot, autostart/silent toggles, start/stop/restart
-  core, quit.
-- **Silent launch**: `--silent` flag or setting — starts to the tray only;
-  first run always shows the window.
-- **Autostart**: `setLoginItemSettings` on Windows/macOS; writes
-  `~/.config/autostart/modelrouter-gui.desktop` on Linux, appending `--silent`
-  when silent mode is on.
 
 ## Layout
 
-```
+```text
-cmd/llmsproxy            # entry point
+cmd/modelrouter          # entry point (pure Go server)
-internal/config          # YAML config loading/validation
+cmd/gui                  # Electron desktop shell (embeds the same core)
-internal/lua             # LuaJIT VM + worker pool + AdapterCache + built-ins (embed)
+internal/config          # config parsing, sealed storage, runtime file
+internal/core            # scheduling, kernel actions, state and audit
+internal/lua             # LuaJIT VM + elastic worker pool + built-in adapters (embed)
+internal/lua/plugins     # optional plugin examples (not shipped in binary/archives)
 internal/provider        # Provider(HTTP) + Registry(routing)
-internal/gateway         # OpenAI-compatible HTTP + auth + SDK/streaming + WebUI
+internal/gateway         # OpenAI-compatible HTTP service + auth + streaming + WebUI
-internal/types           # unified format & OpenAI wire types
+internal/types           # unified format and OpenAI wire types
+packaging/               # deb/rpm/tar packaging, systemd unit, migration script
+docs/                    # topic documentation
 ```
 
-## Development & collaboration
+## Development and tests
-
-- **Git branching workflow**: main / feature / release-branch conventions in
-  **[docs/git-workflow-en.md](docs/git-workflow-en.md)**
-  ([中文](docs/git-workflow.md)). The line: `main` is always deployable →
-  features merge into `main` → cut `release/vX.Y.Z` and tag → release → hotfixes
-  land on the release branch and are cherry-picked back into `main`.
-- **Lua adapter protocol**: see
-  **[docs/lua-adapters-en.md](docs/lua-adapters-en.md)**
-  ([中文](docs/lua-adapters.md)).
-
-## Tests
 
 ```bash
-go test -tags luajit ./...
+go test -tags luajit ./...           # everything
+go test -tags luajit ./internal/lua  # one package (required for LuaJIT changes)
 ```
 
-Covers: config validation, adapter load/transform, signature hooks, gateway
+- Branch and release discipline: [docs/git-workflow-en.md](docs/git-workflow-en.md).
-auth, SDK round trip, SSE streaming, model routing, multimodal passthrough and
+- Plugins live in `plugin_dir`. The gateway never creates, overwrites or deletes
-disable_thinking.
+  plugin files, and a desktop upgrade never touches them either.
+
+## License
+
+MIT License, see [LICENSE](./LICENSE).

cmd/gui/main.js

@@ -1,5 +1,5 @@
 // ModelRouter Desktop GUI — main process
-// Clash-Verge-style: bundles and manages an embedded llmsproxy core instance
+// Desktop shell: bundles and manages an embedded modelrouter core instance
 // in a per-user profile dir (userData), auto-starts it, keeps a tray icon,
 // supports launch-at-login. Server-side users keep using the pure Go binary.
 const {
@@ -18,6 +18,7 @@ const http = require("http");
 const os = require("os");
 const crypto = require("crypto");
 const { spawn, execFileSync } = require("child_process");
+const pluginSync = require("./plugin-sync");
 
 const LOG_FILE = path.join(app.getPath("userData"), "gui.log");
 function log(msg) {
@@ -40,13 +41,36 @@ console.error = function () {
 // Per-user profile dir: <userData>/profile
 const PROFILE_DIR = path.join(app.getPath("userData"), "profile");
 const CONFIG_FILE = path.join(PROFILE_DIR, "config.yaml");
+// Plugin tree for the embedded core. The core deliberately does NOT default
+// plugin_dir (an operator opts in), so the desktop build must set it — without
+// it the whole plugin system stays off and the channels plugin (multi-platform
+// account login + credential vault + auto check-in) is invisible in the WebUI.
+// Seeding is the core's job (SeedBundled fills MISSING files only), so a user's
+// edited copy always wins.
+const PLUGINS_DIR = path.join(PROFILE_DIR, "plugins");
 // Packaged: extraResource at process.resourcesPath/bin (outside asar, spawnable).
 // Dev:   cmd/gui/bin, populated by prepare-core.js.
-const CORE_EXE = path.join(
+// The embedded core is named after the product. Installs produced before the
+// llmsproxy -> modelrouter rename still carry the old file name inside their
+// profile, so resolution falls back to it instead of declaring the core missing.
+const CORE_BIN_DIR = path.join(
   app.isPackaged ? process.resourcesPath : __dirname,
   "bin",
+);
+const CORE_EXE = path.join(
+  CORE_BIN_DIR,
+  process.platform === "win32" ? "modelrouter.exe" : "modelrouter",
+);
+const LEGACY_CORE_EXE = path.join(
+  CORE_BIN_DIR,
   process.platform === "win32" ? "llmsproxy.exe" : "llmsproxy",
 );
+
+function resolveCoreExe() {
+  if (fs.existsSync(CORE_EXE)) return CORE_EXE;
+  if (fs.existsSync(LEGACY_CORE_EXE)) return LEGACY_CORE_EXE;
+  return CORE_EXE;
+}
 // Packaged window icon: must be a REAL file on disk (extraResource), not an
 // asar path — Electron/GTK pulls BrowserWindow icons from the filesystem and
 // crashes (SIGSEGV) when given an asar virtual path on Linux.
@@ -155,7 +179,7 @@ function unsealViaCore() {
   let key = "";
   try {
     const out = execFileSync(
-      CORE_EXE,
+      resolveCoreExe(),
       ["-config", CONFIG_FILE, "-show-secrets"],
       {
         encoding: "utf-8",
@@ -207,7 +231,7 @@ function coreStarted() {
 function writeDefaultEmbeddedConfig() {
   // Generate a self-contained embedded profile with a fresh random admin key.
   // No upstream sources are preconfigured — the desktop user adds them via
-  // the WebUI's sources page (like Clash Verge's empty-by-default profile).
+  // the WebUI's sources page (empty by default, like the WebUI profile).
   // If a config already exists we do NOT overwrite it (keep user's edits).
   if (fs.existsSync(CONFIG_FILE)) return;
   const key = "sk-gw-" + crypto.randomBytes(16).toString("hex");
@@ -216,6 +240,9 @@ function writeDefaultEmbeddedConfig() {
     "listen: 127.0.0.1:" + (settings.port || DEFAULT_PORT),
     "default_model: AUTO",
     "adapter_dir: " + JSON.stringify(path.join(PROFILE_DIR, "adapters")),
+    // Plugin tree: channels/billing/quota/usage ship inside the core binary and
+    // are seeded into this dir on first start. Omitting it disables plugins.
+    "plugin_dir: " + JSON.stringify(PLUGINS_DIR),
     "runtime_file: " + JSON.stringify(path.join(PROFILE_DIR, "runtime.json")),
     "",
     // Admin key as a plain `keys` entry (no seed flag) so the web UI never
@@ -261,15 +288,18 @@ function startCore() {
     console.log("core already running");
     return;
   }
-  if (!fs.existsSync(CORE_EXE)) {
+  const coreExe = resolveCoreExe();
-    console.error("embedded core binary missing:", CORE_EXE);
+  if (!fs.existsSync(coreExe)) {
+    console.error("embedded core binary missing:", coreExe);
     return;
   }
   // ensure profile dir + config + local adapters exist
   fs.mkdirSync(PROFILE_DIR, { recursive: true });
   if (!fs.existsSync(CONFIG_FILE)) writeDefaultEmbeddedConfig();
+  pluginSync.ensurePluginDirConfigured(CONFIG_FILE, PLUGINS_DIR, log);
+  fs.mkdirSync(PLUGINS_DIR, { recursive: true });
 
-  coreProc = spawn(CORE_EXE, ["-config", CONFIG_FILE], {
+  coreProc = spawn(coreExe, ["-config", CONFIG_FILE], {
     cwd: PROFILE_DIR,
     env: Object.assign({}, process.env, {
       LLMS_PROXY_GUI: "1",
@@ -622,12 +652,57 @@ function createWindow() {
     },
   });
   mainWindow.loadFile(path.join(__dirname, "renderer", "index.html"));
+  // External links ALWAYS go to the user's default browser, never an in-app
+  // window. Three paths have to be covered, and missing any one of them is why
+  // "click the login link and nothing happens / the app navigates away" shows up:
+  //   1. window.open / target=_blank            -> setWindowOpenHandler
+  //   2. top-level navigation (same-tab click)  -> will-navigate
+  //   3. navigation inside the WebUI <iframe>   -> will-frame-navigate
+  // The embedded core's own origin (and file:// for the shell) must keep
+  // working, otherwise the WebUI and plugin pages stop loading.
+  const isInternal = (url) =>
+    !url ||
+    url === "about:blank" ||
+    url.startsWith("file://") ||
+    url.startsWith(embeddedBaseUrl()) ||
+    url.startsWith("devtools://");
+  const openExternally = (url) => {
+    if (url && (url.startsWith("http://") || url.startsWith("https://"))) {
+      log("[gui] open in default browser: " + url);
+      shell.openExternal(url);
+    }
+  };
   mainWindow.webContents.setWindowOpenHandler(({ url }) => {
-    if (url && url.startsWith("http")) shell.openExternal(url);
+    if (!isInternal(url)) openExternally(url);
     return { action: "deny" };
   });
+  mainWindow.webContents.on("will-navigate", (e, url) => {
+    if (isInternal(url)) return;
+    e.preventDefault();
+    openExternally(url);
+  });
+  // iframe navigations (the WebUI content area and the plugin pages inside it).
+  // The channels plugin renders its server-side login browser (noVNC) in an
+  // iframe on a LOOPBACK port, so loopback MUST stay in-app — blocking it would
+  // break VNC login. Everything non-loopback is external: open the default
+  // browser instead of silently replacing the gateway UI with a third-party page.
+  const isLoopback = (url) => {
+    try {
+      const h = new URL(url).hostname;
+      return h === "127.0.0.1" || h === "localhost" || h === "[::1]";
+    } catch (e) {
+      return false;
+    }
+  };
+  if (mainWindow.webContents.on) {
+    mainWindow.webContents.on("will-frame-navigate", (e) => {
+      if (isInternal(e.url) || isLoopback(e.url)) return;
+      e.preventDefault();
+      openExternally(e.url);
+    });
+  }
   // Silent start: `--silent` CLI flag or settings.silentStart keeps the app
-  // in the tray without showing the window (Clash-Verge style). First run
+  // in the tray without showing the window (desktop-shell behaviour). First run
   // always shows the window so the user sees the dashboard.
   mainWindow.once("ready-to-show", () => {
     const firstRun = !fs.existsSync(
@@ -751,8 +826,8 @@ ipcMain.handle("core:state", () => ({
   port: settings.port || DEFAULT_PORT,
   profileDir: PROFILE_DIR,
   configFile: CONFIG_FILE,
-  coreExe: CORE_EXE,
+  coreExe: resolveCoreExe(),
-  coreExists: fs.existsSync(CORE_EXE),
+  coreExists: fs.existsSync(resolveCoreExe()),
 }));
 ipcMain.handle("core:start", () => {
   app.isQuitting = false;
@@ -824,7 +899,7 @@ if (gotLock) {
     createWindow();
     createTray();
 
-    // start embedded core if not already running (another instance of llmsproxy)
+    // start embedded core if not already running (another instance of modelrouter)
     setTimeout(async () => {
       const p = await probeEmbedded();
       if (!p.ok) {

cmd/gui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.3",
+  "version": "3.0.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "modelrouter-gui",
-      "version": "1.9.3",
+      "version": "3.0.2",
       "license": "Proprietary",
       "devDependencies": {
         "electron": "^33.0.0",
@@ -1245,7 +1245,7 @@
       }
     },
     "node_modules/chownr": {
-      "version": "3.0.0",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/chownr/-/chownr-3.0.0.tgz",
       "integrity": "sha512-+IxzY9BZOQd/XuYPRmrvEVjF/nqj5kgT4kEq7VofrDoM1MxoRjEWkrCC3EtLi59TVawxTAn+orJwFQcrqEN1+g==",
       "dev": true,
@@ -2288,7 +2288,7 @@
       }
     },
     "node_modules/global-agent": {
-      "version": "3.0.0",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/global-agent/-/global-agent-3.0.0.tgz",
       "integrity": "sha512-PT6XReJ+D07JvGoxQMkT6qji/jVNfX/h364XHZOWeRzy64sSFr+xJ5OX7LI3b4MPQzdL4H8Y8M0xzPpsVMwA8Q==",
       "dev": true,
@@ -2532,7 +2532,7 @@
       "license": "ISC"
     },
     "node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
       "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
       "dev": true,
@@ -2623,7 +2623,7 @@
       }
     },
     "node_modules/json-buffer": {
-      "version": "3.0.1",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/json-buffer/-/json-buffer-3.0.1.tgz",
       "integrity": "sha512-4bV5BfR2mqfQTJm+V5tPPdf+ZpuhiIvTuAB5g8kcrXOZpTT/QwwVRWBywX1ozr6lEuPdbHxwaJlm9G6mI2sfSQ==",
       "dev": true,
@@ -2715,7 +2715,7 @@
       }
     },
     "node_modules/matcher": {
-      "version": "3.0.0",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/matcher/-/matcher-3.0.0.tgz",
       "integrity": "sha512-OkeDaAZ/bQCxeFAozM55PKcKU0yJMPGifLwV4Qgjitu+5MoAfSQN4lsLJeXZ1b8w0x+/Emda6MZgXS1jvsapng==",
       "dev": true,
@@ -3492,7 +3492,7 @@
       }
     },
     "node_modules/shebang-regex": {
-      "version": "3.0.0",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/shebang-regex/-/shebang-regex-3.0.0.tgz",
       "integrity": "sha512-7++dFhtcx3353uBaq8DDR4NuxBetBzC7ZQOhmTQInHEd6bSrXdiEyzCvG07Z44UYdLShWUyXt5M/yhz8ekcb1A==",
       "dev": true,
@@ -3612,7 +3612,7 @@
       }
     },
     "node_modules/sumchecker": {
-      "version": "3.0.1",
+      "version": "3.0.2",
       "resolved": "https://registry.npmjs.org/sumchecker/-/sumchecker-3.0.1.tgz",
       "integrity": "sha512-MvjXzkz/BOfyVDkG0oFOtBxHX2u3gKbMHIF/dXblZsgD3BWOFLmHovIpZY7BykJdAjcqRCBi1WYBNdEC9yI7vg==",
       "dev": true,

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.6",
+  "version": "3.0.2",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",
@@ -39,8 +39,8 @@
       },
       "extraResources": [
         {
-          "from": "bin/llmsproxy",
+          "from": "bin/modelrouter",
-          "to": "bin/llmsproxy"
+          "to": "bin/modelrouter"
         },
         {
           "from": "build/icon.png",
@@ -66,8 +66,8 @@
       "icon": "build/icon.ico",
       "extraResources": [
         {
-          "from": "bin/llmsproxy.exe",
+          "from": "bin/modelrouter.exe",
-          "to": "bin/llmsproxy.exe"
+          "to": "bin/modelrouter.exe"
         },
         {
           "from": "bin/lua51.dll",
@@ -81,6 +81,7 @@
     },
     "files": [
       "main.js",
+      "plugin-sync.js",
       "preload.js",
       "renderer/**/*",
       "build/**/*",
@@ -88,5 +89,5 @@
     ],
     "asar": true
   },
-  "homepage": "https://gitcode.com/JianFeeeee/ModelRouter"
+  "homepage": "https://atomgit.com/HeNanMU-EEA/ModelRouter"
 }

cmd/gui/plugin-sync.js

@@ -0,0 +1,33 @@
+// Plugin-tree maintenance for the desktop profile (no Electron imports, so this
+// is unit-testable with plain node).
+//
+// The desktop may configure an operator-selected plugin directory, but it never
+// owns the files in that directory. In particular, upgrades must not seed,
+// replace, rename, or delete plugins. Plugin installation is an explicit
+// operator action (or an external package manager action).
+const fs = require("fs");
+
+// ensurePluginDirConfigured back-fills `plugin_dir` into an existing config.
+// Appending a top-level YAML key is safe for the shapes this file can have
+// (flat mapping + a keys list). Returns "added" | "present" | "absent" | "error".
+function ensurePluginDirConfigured(configFile, pluginsDir, log) {
+  try {
+    if (!fs.existsSync(configFile)) return "absent";
+    const cur = fs.readFileSync(configFile, "utf-8");
+    if (/^\s*plugin_dir\s*:/m.test(cur)) return "present";
+    fs.writeFileSync(
+      configFile,
+      cur.replace(/\s*$/, "") +
+        "\n# added by ModelRouter Desktop: enables the operator plugin directory\n" +
+        "plugin_dir: " + JSON.stringify(pluginsDir) + "\n",
+      "utf-8",
+    );
+    if (log) log("[gui] plugin_dir back-filled into " + configFile);
+    return "added";
+  } catch (e) {
+    if (log) log("[gui] ensurePluginDirConfigured failed: " + e.message);
+    return "error";
+  }
+}
+
+module.exports = { ensurePluginDirConfigured };

cmd/gui/plugin-sync.test.js

@@ -0,0 +1,48 @@
+// Unit tests for desktop plugin-directory configuration (plain node).
+const test = require("node:test");
+const assert = require("node:assert");
+const fs = require("fs");
+const os = require("os");
+const path = require("path");
+const { ensurePluginDirConfigured } = require("./plugin-sync");
+
+const tmp = () => fs.mkdtempSync(path.join(os.tmpdir(), "plugin-sync-"));
+
+test("back-fills plugin_dir into an existing profile without it", () => {
+  const dir = tmp();
+  const cfg = path.join(dir, "config.yaml");
+  fs.writeFileSync(cfg, "listen: 127.0.0.1:8787\ndefault_model: AUTO\nkeys:\n  - key: k\n    role: admin\n");
+  const plugins = path.join(dir, "plugins");
+  assert.equal(ensurePluginDirConfigured(cfg, plugins, null), "added");
+  const out = fs.readFileSync(cfg, "utf-8");
+  assert.match(out, /^plugin_dir: "/m);
+  assert.match(out, /^listen: 127.0.0.1:8787$/m);
+});
+
+test("leaves a profile that already sets plugin_dir alone", () => {
+  const dir = tmp();
+  const cfg = path.join(dir, "config.yaml");
+  const orig = 'plugin_dir: "/custom/plugins"\nlisten: ":1"\n';
+  fs.writeFileSync(cfg, orig);
+  assert.equal(ensurePluginDirConfigured(cfg, "/whatever", null), "present");
+  assert.equal(fs.readFileSync(cfg, "utf-8"), orig);
+});
+
+test("missing config is not created by the back-fill", () => {
+  const dir = tmp();
+  const cfg = path.join(dir, "config.yaml");
+  assert.equal(ensurePluginDirConfigured(cfg, path.join(dir, "p"), null), "absent");
+  assert.equal(fs.existsSync(cfg), false);
+});
+
+test("desktop upgrade never modifies operator plugin files", () => {
+  const dir = tmp();
+  const plugins = path.join(dir, "plugins");
+  fs.mkdirSync(plugins, { recursive: true });
+  fs.writeFileSync(path.join(plugins, "channels.lua"), "OPERATOR FILE");
+  fs.writeFileSync(path.join(plugins, "my-own.lua"), "MINE");
+  // There is deliberately no version stamp and no upgrade synchronizer.
+  assert.equal(fs.readFileSync(path.join(plugins, "channels.lua"), "utf-8"), "OPERATOR FILE");
+  assert.equal(fs.readFileSync(path.join(plugins, "my-own.lua"), "utf-8"), "MINE");
+  assert.equal(fs.readdirSync(dir).includes(".plugin-seed-version"), false);
+});

cmd/gui/renderer/app.js

@@ -1,5 +1,5 @@
 // ModelRouter Desktop GUI — renderer
-// Clash-Verge-style: the content area is a full-screen iframe loading the
+// Desktop shell: the content area is a full-screen iframe loading the
 // real WebUI 1:1 (status / chat / keys / priority / sources / adapters).
 // The shell only adds the titlebar, settings overlay and tray integration.
 const state = {

cmd/gui/renderer/index.html

@@ -83,7 +83,7 @@
       </div>
     </div>
 
-    <!-- side rail: autostart / silent toggles (Clash-Verge style) -->
+    <!-- side rail: autostart / silent toggles (desktop-shell style) -->
     <aside class="rail" id="rail">
       <div class="rail-gap"></div>
       <button

cmd/gui/renderer/style.css

@@ -1,5 +1,5 @@
 /* =====================================================================
-   ModelRouter Desktop GUI — Clash-Verge-style shell
+   ModelRouter Desktop GUI — desktop-shell-style shell
    Content area = full-screen iframe embedding the real WebUI 1:1.
    Palette mirrors the WebUI (sakura/frost/glassmorphism).
    ===================================================================== */

cmd/gui/scripts/dist-win-docker.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 # One-shot Windows release build for the GUI — ENTIRELY inside docker.
 #
-#   1) cross-compile the embedded core in docker (llmsproxy.exe + lua51.dll)
+#   1) cross-compile the embedded core in docker (modelrouter.exe + lua51.dll)
 #   2) electron-builder NSIS packaging ALSO inside docker (wine for the
 #      uninstaller step lives in the image — see Dockerfile). This is the fix
 #      for the 1.3.0/1.4.0 broken installers: the host had amd64-only wine
@@ -37,13 +37,15 @@ docker run --rm \
   -v "$ROOT":/workspace \
   -v "$CACHE_VOL":/root/.cache \
   -e GOPROXY="$PROXY" \
+  -e ELECTRON_MIRROR="https://npmmirror.com/mirrors/electron/" \
+  -e ELECTRON_BUILDER_BINARIES_MIRROR="https://npmmirror.com/mirrors/electron-builder-binaries/" \
   -w /workspace \
   "$IMAGE" bash -c '
     set -euo pipefail
     cp /opt/luajit2/src/lua51.dll /workspace/cmd/gui/bin/lua51.dll
     GOOS=windows GOARCH=amd64 CGO_ENABLED=1 CC=x86_64-w64-mingw32-gcc \
-      go build -buildvcs=false -tags luajit -trimpath \
+      go build -buildvcs=false -tags luajit -trimpath -ldflags "-s -w" \
-      -o /workspace/cmd/gui/bin/llmsproxy.exe ./cmd/llmsproxy
+      -o /workspace/cmd/gui/bin/modelrouter.exe ./cmd/modelrouter
     echo "  -> core built"
     # wine needs one prefix initialization so its wow64 runtime (syswow64/ntdll)
     # exists — the exact thing that was missing on the broken host. Use a shared

cmd/gui/scripts/prepare-core.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-// Build the embedded llmsproxy core(s) for the GUI package.
+// Build the embedded modelrouter core(s) for the GUI package.
 //   node prepare-core.js        -> linux (native, luajit) + win (cross, best effort)
 //   node prepare-core.js linux  -> linux only
 //   node prepare-core.js win    -> win only
@@ -12,21 +12,24 @@ const binDir = path.join(guiDir, "bin");
 const projectRoot = path.join(guiDir, "..", "..");
 
 function buildLinux() {
-  const out = path.join(binDir, "llmsproxy");
+  const out = path.join(binDir, "modelrouter");
-  console.log("[prepare-core] building llmsproxy (linux/luajit)...");
+  console.log("[prepare-core] building modelrouter (linux/luajit)...");
   fs.mkdirSync(binDir, { recursive: true });
   execFileSync(
     "go",
-    ["build", "-tags", "luajit", "-trimpath", "-o", out, "./cmd/llmsproxy"],
+    // -s -w: same stripping as packaging/core-dist.sh. Without it the bundled
+    // core carries debug info (~14.4 MB vs ~10.3 MB) and every installer pays
+    // for it, for a binary the user never debugs.
+    ["build", "-tags", "luajit", "-trimpath", "-ldflags", "-s -w", "-o", out, "./cmd/modelrouter"],
     { cwd: projectRoot, stdio: "inherit" },
   );
   return out;
 }
 
 function buildWin() {
-  const out = path.join(binDir, "llmsproxy.exe");
+  const out = path.join(binDir, "modelrouter.exe");
   console.log(
-    "[prepare-core] cross-building llmsproxy.exe (windows/amd64, luajit)...",
+    "[prepare-core] cross-building modelrouter.exe (windows/amd64, luajit)...",
   );
   fs.mkdirSync(binDir, { recursive: true });
   const cc = process.env.MINGW_CC || "x86_64-w64-mingw32-gcc";
@@ -43,7 +46,7 @@ function buildWin() {
   }
   execFileSync(
     "go",
-    ["build", "-tags", "luajit", "-trimpath", "-o", out, "./cmd/llmsproxy"],
+    ["build", "-tags", "luajit", "-trimpath", "-ldflags", "-s -w", "-o", out, "./cmd/modelrouter"],
     {
       cwd: projectRoot,
       stdio: "inherit",

cmd/modelrouter/main.go

@@ -1,4 +1,4 @@
-// Command llmsproxy is a standalone gateway that exposes multiple upstream LLM
+// Command modelrouter is a standalone gateway that exposes multiple upstream LLM
 // sources behind a unified OpenAI-compatible HTTP API. Protocol differences are
 // handled by Lua adapters (per source), optionally signing outgoing requests.
 // It supports explicit model selection or AUTO routing by priority, image
@@ -16,9 +16,9 @@ import (
 	"syscall"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
-	"llmsproxy/internal/gateway"
+	"modelrouter/internal/gateway"
 )
 
 func main() {
@@ -33,12 +33,12 @@ func main() {
 	// error here instead of being silently created.
 	if *checkOnly {
 		if _, err := os.Stat(*cfgPath); err != nil {
-			log.Fatalf("[llmsproxy] check: %v", err)
+			log.Fatalf("[modelrouter] check: %v", err)
 		}
 		if _, err := config.Load(*cfgPath); err != nil {
-			log.Fatalf("[llmsproxy] check: %v", err)
+			log.Fatalf("[modelrouter] check: %v", err)
 		}
-		log.Printf("[llmsproxy] check: %s is valid", *cfgPath)
+		log.Printf("[modelrouter] check: %s is valid", *cfgPath)
 		return
 	}
 
@@ -48,34 +48,34 @@ func main() {
 	// program prints a credential, and this path never logs to a file.
 	if *showSecrets {
 		if _, err := os.Stat(*cfgPath); err != nil {
-			log.Fatalf("[llmsproxy] show-secrets: %v", err)
+			log.Fatalf("[modelrouter] show-secrets: %v", err)
 		}
 		if err := config.PrintSecrets(*cfgPath); err != nil {
-			log.Fatalf("[llmsproxy] show-secrets: %v", err)
+			log.Fatalf("[modelrouter] show-secrets: %v", err)
 		}
 		return
 	}
 
 	created, err := config.EnsureDefault(*cfgPath)
 	if err != nil {
-		log.Fatalf("[llmsproxy] config: %v", err)
+		log.Fatalf("[modelrouter] config: %v", err)
 	}
 
 	c, err := core.New(*cfgPath)
 	if err != nil {
-		log.Fatalf("[llmsproxy] core: %v", err)
+		log.Fatalf("[modelrouter] core: %v", err)
 	}
 	defer c.Close()
 
 	if created {
 		for _, k := range c.GatewayKeys() {
-			log.Printf("[llmsproxy] generated default config at %s — admin key: %s (rotate it in the WebUI after first login)", *cfgPath, k)
+			log.Printf("[modelrouter] generated default config at %s — admin key: %s (rotate it in the WebUI after first login)", *cfgPath, k)
 		}
 	}
 
 	gw, err := gateway.New(c)
 	if err != nil {
-		log.Fatalf("[llmsproxy] gateway: %v", err)
+		log.Fatalf("[modelrouter] gateway: %v", err)
 	}
 
 	// Ops hygiene: surface the two most common footguns instead of silently
@@ -88,16 +88,16 @@ func main() {
 	// (e.g. after rotating the starter key away) it reported "every request is
 	// rejected" while seven working keys, one of them admin, were in service.
 	if keys := c.ListKeys(); len(keys) == 0 {
-		log.Printf("[llmsproxy] WARNING: no gateway key configured — every request will be rejected. Add one in the WebUI or set gateway_keys")
+		log.Printf("[modelrouter] WARNING: no gateway key configured — every request will be rejected. Add one in the WebUI or set gateway_keys")
 	} else {
 		for _, k := range keys {
 			if k.Key == "sk-gw-local-0001" || k.Key == "sk-local-0001" {
-				log.Printf("[llmsproxy] WARNING: gateway key %q looks like the starter/example key — rotate it before exposing the gateway", k.Key)
+				log.Printf("[modelrouter] WARNING: gateway key %q looks like the starter/example key — rotate it before exposing the gateway", k.Key)
 			}
 		}
 	}
 	if l := c.Listen(); strings.HasPrefix(l, "0.0.0.0:") || strings.HasPrefix(l, "::") {
-		log.Printf("[llmsproxy] WARNING: listen=%s binds ALL interfaces — bind an internal address in production", l)
+		log.Printf("[modelrouter] WARNING: listen=%s binds ALL interfaces — bind an internal address in production", l)
 	}
 
 	srv := &http.Server{
@@ -108,22 +108,22 @@ func main() {
 	go func() {
 		cert, key := c.TLS()
 		if cert != "" && key != "" {
-			log.Printf("[llmsproxy] listening HTTPS on %s (cert=%q) (default_model=%s, models=%v, adapters=%d)",
+			log.Printf("[modelrouter] listening HTTPS on %s (cert=%q) (default_model=%s, models=%v, adapters=%d)",
 				c.Listen(), cert, c.DefaultModel(), c.Registry().ModelList(), len(c.ListAdapters()))
 			if err := srv.ListenAndServeTLS(cert, key); err != nil && err != http.ErrServerClosed {
-				log.Fatalf("[llmsproxy] server: %v", err)
+				log.Fatalf("[modelrouter] server: %v", err)
 			}
 			return
 		}
-		log.Printf("[llmsproxy] listening HTTP on %s (default_model=%s, models=%v, adapters=%d)",
+		log.Printf("[modelrouter] listening HTTP on %s (default_model=%s, models=%v, adapters=%d)",
 			c.Listen(), c.DefaultModel(), c.Registry().ModelList(), len(c.ListAdapters()))
 		if err := srv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
-			log.Fatalf("[llmsproxy] server: %v", err)
+			log.Fatalf("[modelrouter] server: %v", err)
 		}
 	}()
 
 	stop := make(chan os.Signal, 1)
 	signal.Notify(stop, syscall.SIGINT, syscall.SIGTERM)
 	<-stop
-	log.Printf("[llmsproxy] shutting down")
+	log.Printf("[modelrouter] shutting down")
 }

deploy.sh

@@ -1,10 +1,10 @@
 #!/usr/bin/env bash
 # ============================================================
-# llmsproxy 原子化滚动部署脚本
+# modelrouter 原子化滚动部署脚本
 #
 # 策略：
 #   1. /tmp (tmpfs) 构建，/usr/local/bin (ext4) 部署
-#   2. 构建后先 copy 到目标文件系统（.llmsproxy.tmp），
+#   2. 构建后先 copy 到目标文件系统（.modelrouter.tmp），
 #      再 rename(2) 做原子替换（同一文件系统内 mv 是原子的）
 #   3. runtime.json 不替换（保留运行时密钥、审计状态）
 #   4. systemctl restart → 服务中断仅在 Go 进程重启瞬间 (<500ms)
@@ -28,15 +28,24 @@
 set -euo pipefail
 
 # ---------- 配置 ----------
-readonly TARGET_BIN="/usr/local/bin/llmsproxy"
+readonly TARGET_BIN="/usr/local/bin/modelrouter"
-readonly TARGET_CONFIG="/etc/llmsproxy/config.yaml"
+readonly TARGET_CONFIG="/etc/modelrouter/config.yaml"
-readonly TARGET_ADAPTERS="/etc/llmsproxy/adapters"
+readonly TARGET_ADAPTERS="/etc/modelrouter/adapters"
-readonly REPO_DIR="/home/program/llmsproxy"
+readonly TARGET_PLUGINS="/etc/modelrouter/plugins"
+# 仓库所在目录不等于产品名：这里保持真实路径，除非用 MODELROUTER_REPO_DIR 覆盖。
+readonly REPO_DIR="${MODELROUTER_REPO_DIR:-/home/program/llmsproxy}"
 readonly HEALTH_URL="http://127.0.0.1:8081/v1/models"
-readonly BIN_NAME="llmsproxy"
+readonly BIN_NAME="modelrouter"
 
-# 构建目录在 /tmp（tmpfs 更快）
+# 构建目录。默认落在仓库下的构建工作区，而不是 /tmp：本机 /tmp 是 8.8G
-readonly TMP_DIR="/tmp/llmsproxy-deploy"
+# tmpfs，实测被 Go 构建缓存占满后会让 go test/build 在链接阶段直接失败。
+# 可用 MODELROUTER_BUILD_DIR 覆盖；同一目录也被测试用作 GOTMPDIR。
+readonly BUILD_SUBDIR=".build-work"
+if [[ -n "${MODELROUTER_BUILD_DIR:-}" ]]; then
+    readonly TMP_DIR="${MODELROUTER_BUILD_DIR}"
+else
+    readonly TMP_DIR="${REPO_DIR}/${BUILD_SUBDIR}"
+fi
 # 同文件系统暂存文件（用于原子 rename）
 readonly STAGING_BIN="${TARGET_BIN}.staging"
 readonly STAGING_CONFIG="${TARGET_CONFIG}.staging"
@@ -51,7 +60,7 @@ usage() {
     cat <<'EOF'
 用法: ./deploy.sh [--config <file>]
 
-  --config <file>   投放这份配置到 /etc/llmsproxy/config.yaml（原子替换）。
+  --config <file>   投放这份配置到 /etc/modelrouter/config.yaml（原子替换）。
                     省略时不改动线上配置，只部署二进制与适配器。
                     无论是否指定，重启前都会用新二进制 -check 预检线上配置。
 
@@ -95,6 +104,34 @@ precheck() {
     command -v systemctl &>/dev/null || fail "未找到 systemctl"
     [[ $(id -u) -eq 0 ]] || fail "需要 root 权限"
 
+    # 改名保护：本脚本现在部署到 /etc/modelrouter。若机器上仍是改名前的
+    # /etc/llmsproxy 且新路径尚未建立，直接继续会创建一份**空配置**并让新
+    # 服务在没有 key/源的情况下起来 —— 看起来成功，实际网关拒掉所有请求。
+    if [[ -d /etc/llmsproxy && ! -e "$TARGET_CONFIG" ]]; then
+        fail "检测到改名前的部署 /etc/llmsproxy，但 $TARGET_CONFIG 不存在。
+     先迁移（dry-run 默认，确认后加 --apply）：
+       sudo ./packaging/migrate-to-modelrouter.sh
+       sudo ./packaging/migrate-to-modelrouter.sh --apply
+     如需全新部署，先把旧目录移开：mv /etc/llmsproxy /etc/llmsproxy.bak"
+    fi
+
+    # unit 与二进制路径必须一致。包安装把二进制放 /usr/bin，源码部署放
+    # /usr/local/bin；unit 里写错一行，systemctl restart 只会给 203/EXEC，
+    # 看起来像"部署成功但服务起不来"。此处提前拦住。
+    local unit="/etc/systemd/system/modelrouter.service"
+    if [[ -f "$unit" ]]; then
+        local unit_exe
+        unit_exe=$(awk -F'=' '/^ExecStart=/{print $2; exit}' "$unit" | awk '{print $1}')
+        if [[ -n "$unit_exe" && "$unit_exe" != "$TARGET_BIN" && ! -x "$unit_exe" ]]; then
+            fail "unit 的 ExecStart 指向 $unit_exe，但该文件不存在。
+     包安装（deb/rpm）用 /usr/bin/modelrouter；源码部署（本脚本）用 $TARGET_BIN。
+     请把 $unit 的 ExecStart 改成 $TARGET_BIN，或改用包安装。"
+        fi
+        if [[ -n "$unit_exe" && "$unit_exe" != "$TARGET_BIN" ]]; then
+            warn "unit ExecStart=$unit_exe，而本脚本部署到 $TARGET_BIN（两者都存在，确认哪个该生效）"
+        fi
+    fi
+
     if [[ -n "$NEW_CONFIG_SRC" ]]; then
         [[ -f "$NEW_CONFIG_SRC" ]] || fail "待部署配置不存在: $NEW_CONFIG_SRC"
         # 同一文件比对：--config 指向线上文件本身时没有"投放"动作可做，
@@ -121,14 +158,14 @@ backup_old_state() {
     log "备份旧二进制到 $BACKUP_BIN"
     if [[ -f "$TARGET_BIN" ]]; then
         cp -f "$TARGET_BIN" "$BACKUP_BIN"
-        log "✓ 旧二进制已备份 (sha256=$(sha256sum "$BACKUP_BIN" | awk '{print $1}' | cut -c1-16))"
+        log "OK 旧二进制已备份 (sha256=$(sha256sum "$BACKUP_BIN" | awk '{print $1}' | cut -c1-16))"
     else
         warn "旧二进制不存在，跳过备份（无回滚点）"
     fi
 
     if [[ -f "$TARGET_CONFIG" ]]; then
         cp -f "$TARGET_CONFIG" "$BACKUP_CONFIG"
-        log "✓ 旧配置已备份到 $BACKUP_CONFIG (sha256=$(sha256sum "$BACKUP_CONFIG" | awk '{print $1}' | cut -c1-16))"
+        log "OK 旧配置已备份到 $BACKUP_CONFIG (sha256=$(sha256sum "$BACKUP_CONFIG" | awk '{print $1}' | cut -c1-16))"
     else
         warn "线上配置不存在: $TARGET_CONFIG（首次部署？无配置回滚点）"
     fi
@@ -141,11 +178,11 @@ build() {
     mkdir -p "$TMP_DIR/bin"
 
     cd "$REPO_DIR"
-    CGO_ENABLED=1 go build -tags luajit \
+    CGO_ENABLED=1 GOTMPDIR="$TMP_DIR" go build -tags luajit \
         -trimpath \
         -ldflags="-s -w" \
         -o "$TMP_DIR/bin/$BIN_NAME" \
-        ./cmd/llmsproxy
+        ./cmd/modelrouter
 
     NEW_BIN="$TMP_DIR/bin/$BIN_NAME"
     NEW_SIZE=$(stat -c%s "$NEW_BIN")
@@ -174,9 +211,9 @@ build() {
     )
     for c in "${checks[@]}"; do
         if strings "$NEW_BIN" | grep -qF "$c"; then
-            log "  ✓ 新二进制包含: $c"
+            log "  OK 新二进制包含: $c"
         else
-            warn "  ⚠ 新二进制不包含: $c（非预期但非致命）"
+            warn "  注意： 新二进制不包含: $c（非预期但非致命）"
         fi
     done
 }
@@ -206,7 +243,7 @@ atomic_replace_binary() {
 
     # 验证
     NEW_SHA256=$(sha256sum "$TARGET_BIN" | awk '{print $1}')
-    log "✓ 二进制已原子替换 (sha256=${NEW_SHA256:0:16})"
+    log "OK 二进制已原子替换 (sha256=${NEW_SHA256:0:16})"
 }
 
 # ---------- 原子替换配置 ----------
@@ -222,7 +259,7 @@ atomic_replace_config() {
 
     log "  rename → $TARGET_CONFIG (原子)"
     mv -f "$STAGING_CONFIG" "$TARGET_CONFIG"
-    log "✓ 配置已原子替换 (sha256=$(sha256sum "$TARGET_CONFIG" | awk '{print $1}' | cut -c1-16))"
+    log "OK 配置已原子替换 (sha256=$(sha256sum "$TARGET_CONFIG" | awk '{print $1}' | cut -c1-16))"
 }
 
 # ---------- 配置预检 ----------
@@ -235,7 +272,7 @@ verify_config() {
 
     local out
     if out=$("$TARGET_BIN" -check -config "$TARGET_CONFIG" 2>&1); then
-        log "✓ 配置合法"
+        log "OK 配置合法"
         return 0
     fi
 
@@ -246,7 +283,7 @@ verify_config() {
 
 # ---------- 密钥可解封校验 ----------
 # config.yaml 里的凭据以 enc:v1: 密文落盘，解密钥匙是 master.key（或
-# LLMS_PROXY_MASTER_KEY）。一旦钥匙丢失/错位，配置在**启动时**才会报错，
+# MODELROUTER_MASTER_KEY）。一旦钥匙丢失/错位，配置在**启动时**才会报错，
 # 那时服务已经处于「二进制已替换 + 配置已替换」的半死状态，只能靠回滚。
 # 所以在替换任何东西之前先验：能不能用现有的钥匙解开现有的密文。
 #
@@ -271,27 +308,28 @@ verify_master_key() {
     [[ -n "$runtime_file" ]] || runtime_file="$(dirname "$TARGET_CONFIG")/runtime.json"
     key_file="$(dirname "$runtime_file")/master.key"
 
-    if [[ -n "${LLMS_PROXY_MASTER_KEY:-}" ]]; then
+    MASTER_KEY_ENV="${MODELROUTER_MASTER_KEY:-${LLMS_PROXY_MASTER_KEY:-}}"
-        log "  钥匙来源: LLMS_PROXY_MASTER_KEY（环境变量）"
+    if [[ -n "$MASTER_KEY_ENV" ]]; then
+        log "  钥匙来源: MODELROUTER_MASTER_KEY（环境变量）"
         # 走真实解密路径验证，而不是只检查格式
         if out=$("$TARGET_BIN" -show-secrets -config "$TARGET_CONFIG" 2>&1); then
-            log "✓ 凭据可解封（$(echo "$out" | grep -cE '^(source|key) ') 项）"
+            log "OK 凭据可解封（$(echo "$out" | grep -cE '^(source|key) ') 项）"
             return 0
         fi
-        err "凭据无法用 LLMS_PROXY_MASTER_KEY 解封，未替换任何文件："
+        err "凭据无法用 MODELROUTER_MASTER_KEY 解封，未替换任何文件："
         printf '%s\n' "$out" >&2
         return 1
     fi
 
     if [[ ! -f "$key_file" ]]; then
         err "配置含 $sealed 处密文，但钥匙文件不存在: $key_file"
-        err "  没有它，部署后服务将无法启动。恢复 master.key 或设置 LLMS_PROXY_MASTER_KEY 后重试。"
+        err "  没有它，部署后服务将无法启动。恢复 master.key 或设置 MODELROUTER_MASTER_KEY 后重试。"
         return 1
     fi
     log "  钥匙文件: $key_file ($(stat -c '%A' "$key_file" 2>/dev/null || echo '?')，$(wc -c < "$key_file") 字节)"
 
     if out=$("$TARGET_BIN" -show-secrets -config "$TARGET_CONFIG" 2>&1); then
-        log "✓ 凭据可解封（$(echo "$out" | grep -cE '^(source|key) ') 项）"
+        log "OK 凭据可解封（$(echo "$out" | grep -cE '^(source|key) ') 项）"
         return 0
     fi
     err "凭据无法解密（钥匙与配置不匹配），未替换任何文件："
@@ -301,48 +339,60 @@ verify_master_key() {
     return 1
 }
 
-# ---------- 适配器备份保留策略 ----------
+# ---------- 备份目录保留策略 ----------
-# 每次部署都新建一个 adapters.bak.<时间戳> 目录，而回滚只用到最近一次
+# 每次部署都新建一个 <prefix>.bak.<时间戳> 目录，而回滚只用到最近一次
 # （$BACKUP_BIN / $BACKUP_CONFIG 都是单文件覆盖）。多出来的目录从没人清理，
-# 实测线上已积累 71 个，/etc/llmsproxy 因此涨到 122M。
+# 实测线上已积累 71 个，/etc/modelrouter 因此涨到 122M。
 #
-# 保留最近 KEEP_ADAPTER_BACKUPS 份足够回滚，同时给目录数设上限——否则
+# 保留最近 KEEP_BACKUPS 份足够回滚，同时给目录数设上限——否则一次误配置
-# 一次误配置（比如 adapter_dir 指错）就可能在几秒内造出成百上千个目录。
+# 就可能在几秒内造出成百上千个目录。
-# 只删名字严格匹配 adapters.bak.<14位时间戳> 的目录，避免误伤人工放的目录。
+# 只删名字严格匹配 <prefix>.bak.<14位时间戳> 的目录，避免误伤人工放的目录。
-KEEP_ADAPTER_BACKUPS=5
+#
+# ★ 必须**同时覆盖 adapters 与 plugins**（同一缺陷的两个实例）：
+#   本函数最初只处理 adapters.bak，而 sync_plugins 同样每次建 plugins.bak.*
+#   —— 于是修了 adapters 的堆积、plugins 继续涨（实测线上 36 个、占用 192M）。
+#   这正是「两处只修一处」的老形状：给一个新前缀加备份目录时，必须同步纳入
+#   prune_all_backups，否则它就是一个不会被清理的新泄漏点。
+KEEP_BACKUPS=5
 
-prune_adapter_backups() {
+prune_backups() {
-    local base="/etc/llmsproxy"
+    local prefix="$1"
+    local base="/etc/modelrouter"
     [[ -d "$base" ]] || return 0
+    # 防误用：前缀必须是已知的两个（避免有人传空串删掉整个目录）
+    case "$prefix" in
+        adapters|plugins) ;;
+        *) fail "prune_backups: 未知前缀 '$prefix'（只接受 adapters|plugins）" ;;
+    esac
 
     # 先按数量上限硬裁：即使时间戳排序失效也不会无上限增长。
     local all
-    mapfile -t all < <(find "$base" -maxdepth 1 -type d -name 'adapters.bak.*' -printf '%f\n' | sort)
+    mapfile -t all < <(find "$base" -maxdepth 1 -type d -name "$prefix.bak.*" -printf '%f\n' | sort)
-    local cap=$((KEEP_ADAPTER_BACKUPS * 4))
+    local cap=$((KEEP_BACKUPS * 4))
     if (( ${#all[@]} > cap )); then
-        warn "适配器备份目录有 ${#all[@]} 个（异常），裁到 $cap"
+        warn "$prefix 备份目录有 ${#all[@]} 个（异常），裁到 $cap"
         local i=0
         for d in "${all[@]}"; do
             i=$((i + 1))
             # 从最旧的开始删（sort 后升序）。名字不规范的跳过不删。
             if (( i <= ${#all[@]} - cap )); then
-                if [[ "$d" =~ ^adapters\.bak\.[0-9]{14}$ ]]; then
+                if [[ "$d" =~ ^"$prefix"\.bak\.[0-9]{14}$ ]]; then
                     rm -rf "${base:?}/$d"
                 else
                     warn "跳过名字不规范的备份目录（不删）: $d"
                 fi
             fi
         done
-        mapfile -t all < <(find "$base" -maxdepth 1 -type d -name 'adapters.bak.*' -printf '%f\n' | sort)
+        mapfile -t all < <(find "$base" -maxdepth 1 -type d -name "$prefix.bak.*" -printf '%f\n' | sort)
     fi
 
-    # 再按时间保留最近 KEEP_ADAPTER_BACKUPS 份（sort 后最新在末尾）。
+    # 再按时间保留最近 KEEP_BACKUPS 份（sort 后最新在末尾）。
     local total=${#all[@]} i=0
     for d in "${all[@]}"; do
         i=$((i + 1))
         # i <= total-KEEP 的是较旧的，要删。
-        if (( i <= total - KEEP_ADAPTER_BACKUPS )); then
+        if (( i <= total - KEEP_BACKUPS )); then
-            if [[ "$d" =~ ^adapters\.bak\.[0-9]{14}$ ]]; then
+            if [[ "$d" =~ ^"$prefix"\.bak\.[0-9]{14}$ ]]; then
                 rm -rf "${base:?}/$d"
             else
                 warn "跳过名字不规范的备份目录（不删）: $d"
@@ -351,8 +401,15 @@ prune_adapter_backups() {
     done
 
     local left
-    left=$(find "$base" -maxdepth 1 -type d -name 'adapters.bak.*' | wc -l)
+    left=$(find "$base" -maxdepth 1 -type d -name "$prefix.bak.*" | wc -l)
-    log "  适配器备份保留最近 $KEEP_ADAPTER_BACKUPS 份（当前剩 $left 个）"
+    log "  $prefix 备份保留最近 $KEEP_BACKUPS 份（当前剩 $left 个）"
+}
+
+# prune_all_backups：两个前缀都裁。
+# ★ 新增任何 .bak.<时间戳> 目录的前缀时，必须在这里补上 —— 否则又一个泄漏点。
+prune_all_backups() {
+    prune_backups adapters
+    prune_backups plugins
 }
 
 # ---------- 同步适配器 ----------
@@ -366,35 +423,42 @@ sync_adapters() {
     [[ -d "$SRC_ADAPTERS" ]] || fail "内嵌适配器目录不存在: $SRC_ADAPTERS"
 
     # 备份旧的适配器文件（含任何运行时覆盖版本，如 opencode.lua）
-    BACKUP_DIR="/etc/llmsproxy/adapters.bak.$(date +%Y%m%d%H%M%S)"
+    BACKUP_DIR="/etc/modelrouter/adapters.bak.$(date +%Y%m%d%H%M%S)"
     if [[ -d "$TARGET_ADAPTERS" ]]; then
         mkdir -p "$BACKUP_DIR"
         cp -rf "$TARGET_ADAPTERS/"*.lua "$BACKUP_DIR/" 2>/dev/null || true
         log "  旧适配器已备份到 $BACKUP_DIR"
     fi
-    prune_adapter_backups
+    prune_all_backups
 
     cp -f "$SRC_ADAPTERS/"*.lua "$TARGET_ADAPTERS/"
     chmod 0644 "$TARGET_ADAPTERS/"*.lua
 
     # 校验：每个适配器必须包含 usage 透传修复（opencode/openai/deepseek 等应有 'uses'）
-    local required_files=(openai deepseek anthropic gemini ollama opencode github groq kimicode mistral sensenova agentrouter)
+    local required_files=(openai deepseek anthropic gemini ollama github groq mistral sensenova qwen38)
     for f in "${required_files[@]}"; do
         [[ -f "$TARGET_ADAPTERS/$f.lua" ]] || warn "  缺少适配器: $f.lua"
     done
     local usage_fixed=0
-    for f in openai deepseek github groq kimicode mistral opencode; do
+    for f in openai deepseek github groq mistral; do
         grep -qE '\buses\b' "$TARGET_ADAPTERS/$f.lua" 2>/dev/null && ((usage_fixed++)) || true
     done
-    log "  usage 透传修复适配器: ${usage_fixed}/7 (openai/deepseek/github/groq/kimicode/mistral/opencode)"
+    log "  usage 透传修复适配器: ${usage_fixed}/5 (openai/deepseek/github/groq/mistral)"
-    log "✓ 适配器已同步到 $TARGET_ADAPTERS/"
+    log "OK 适配器已同步到 $TARGET_ADAPTERS/"
 }
 
+# 插件不在这里同步。
+#
+# 早期实现会把内置插件复制/覆盖到 plugin_dir，理由是"内置插件是发行物的一部分"。
+# 该前提已被废止：插件不再编入二进制、不进发行包、不自动安装，属于运维自有资产。
+# 部署脚本只负责适配器（协议转换，随内核发布）与二进制/配置，绝不创建、覆盖或
+# 删除 plugin_dir 里的任何文件。
+
 # ---------- 重启服务 ----------
 restart_service() {
     log "重启服务（中断 <500ms）"
-    systemctl restart llmsproxy.service || fail "systemd restart 失败"
+    systemctl restart modelrouter.service || fail "systemd restart 失败"
-    log "✓ llmsproxy.service restarted"
+    log "OK modelrouter.service restarted"
 }
 
 # ---------- 健康检查 ----------
@@ -409,7 +473,7 @@ healthcheck() {
         local code
         code=$(curl -s -o /dev/null -w "%{http_code}" --max-time 3 "$HEALTH_URL" 2>/dev/null || echo "000")
         if [[ "$code" == "200" || "$code" == "401" ]]; then
-            log "✓ 服务健康 (status=$code, ${attempt}/${attempts})"
+            log "OK 服务健康 (status=$code, ${attempt}/${attempts})"
             return 0
         fi
         log "  等待中 ($attempt/$attempts, status=$code)"
@@ -417,7 +481,7 @@ healthcheck() {
         (( attempt++ ))
     done
 
-    warn "健康检查未通过，请检查服务状态: systemctl status llmsproxy.service"
+    warn "健康检查未通过，请检查服务状态: systemctl status modelrouter.service"
     return 1
 }
 
@@ -453,7 +517,7 @@ rollback() {
 
     if (( restored_any == 0 )); then
         err "无任何回滚点，无法自动恢复"
-        err "请手动检查: systemctl status llmsproxy.service && journalctl -u llmsproxy -n 50"
+        err "请手动检查: systemctl status modelrouter.service && journalctl -u modelrouter -n 50"
         return 1
     fi
 
@@ -467,9 +531,9 @@ rollback() {
     fi
 
     log "重启服务（旧版本）"
-    systemctl restart llmsproxy.service || {
+    systemctl restart modelrouter.service || {
         err "重启失败，服务可能完全宕机"
-        err "请手动检查: systemctl status llmsproxy.service"
+        err "请手动检查: systemctl status modelrouter.service"
         return 1
     }
 
@@ -478,7 +542,7 @@ rollback() {
     local code
     code=$(curl -s -o /dev/null -w "%{http_code}" --max-time 5 "$HEALTH_URL" 2>/dev/null || echo "000")
     if [[ "$code" == "200" || "$code" == "401" ]]; then
-        log "✓ 回滚成功，旧版本已恢复服务 (status=$code)"
+        log "OK 回滚成功，旧版本已恢复服务 (status=$code)"
         log "  新版本部署失败，请检查构建/配置后重试"
         return 0
     else
@@ -495,6 +559,7 @@ print_summary() {
     log "  二进制: $TARGET_BIN"
     log "  配置:   $TARGET_CONFIG$([[ -n "$NEW_CONFIG_SRC" ]] && echo "  (本次已替换)" || echo "  (未改动)")"
     log "  适配器: $TARGET_ADAPTERS"
+    log "  插件:   $TARGET_PLUGINS"
     log "═════════════════════════════════════════════"
 }
 
@@ -503,7 +568,7 @@ print_summary() {
 # 于是坏配置在服务仍然健康时就被拦下，根本不会触发重启。
 main() {
     parse_args "$@"
-    log "═══ llmsproxy 原子部署开始 ═══"
+    log "═══ modelrouter 原子部署开始 ═══"
     precheck
     # 密钥校验放在 build/替换之前：它的失败后果最重（配置里的密文永久解不开），
     # 必须在动任何文件之前就挡住。此阶段用的是**当前线上二进制**，它已经

docs/adapters.md

@@ -0,0 +1,61 @@
+# 适配器协议
+
+> 本文档由 README 拆分而来：Lua 适配器协议与内置适配器清单。
+> 编写适配器的完整参考见 [lua-adapters.md](lua-adapters.md)。
+> 返回 [README](../README.md)。
+
+## Lua 适配器协议
+
+完整 API 见 **[Lua 适配器 API 文档](docs/lua-adapters.md)**（[English](docs/lua-adapters-en.md)）。
+
+每个适配器是一个返回 table 的 Lua 脚本（`internal/lua/adapters/<name>.lua`），
+加载时可被 `adapter_dir` 下的同名脚本覆盖——**改 Lua 脚本无需重编译**，
+重启即生效；通过 WebUI 上传的适配器与在线编辑的源配置则即时生效。
+
+```lua
+return {
+  name = "mysrc",
+  version = "1.0.0",
+  endpoint = "/chat/completions",        -- 上游路径（可被 source.endpoint 覆盖）
+  headers = { ["X-Static"] = "v" },      -- 静态默认请求头（build_headers 缺省时使用）
+
+  -- 请求转换：把 OpenAI 格式 req 转成上游原生格式，返回字符串
+  transform_request = function(raw_json) ... end,
+
+  -- 响应转换：把上游原生响应转成统一格式字符串
+  --   { content, reasoning_content, finish_reason, token_usage{...}, tool_calls[{...}] }
+  transform_response = function(raw_json) ... end,
+
+  -- 流式分块转换：把上游 SSE data 转成 { content, done, ... }，返回 "" 则跳过
+  transform_stream_chunk = function(raw_chunk) ... end,
+
+  -- [可选] 动态请求头/签名钩子
+  -- meta = { url, method, body, api_key, timestamp, source={ name, meta={...} } }
+  build_headers = function(meta) return { ["X-App-Sign"] = sign } end,
+
+  -- [可选] 出向序列化（契约 v2）：上游不是 OpenAI 协议时接管写给客户端的字节。
+  -- 不声明则沿用内核的 OpenAI 兜底；声明即完全取代，内核不再混合字段。
+  serialize_response = function(result) return body_str end,
+  serialize_open_preamble = function(result) return bytes end,
+  serialize_chunk = function(delta) return bytes end,
+  serialize_close = function(result) return bytes end,
+}
+```
+
+内置信号辅助：`hmac_sha256_hex(key, data)`、`sha256_hex(data)`、`base64_encode(s)`、
+`tohex(s)`、`json.encode/decode`、`log(level, msg)`。
+
+### 内置适配器
+
+`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `qwen38` `sensenova`
+`qwen38` `sensenova` `trae`。
+
+`anthropic`/`gemini`/`ollama` 适配器内置多模态转换（`image_url` → 各自上游格式）；
+推理开关由中立的 `thinking` 字段表达，`deepseek` 适配器把它映射为
+`thinking.type = "disabled"`。
+
+适配器可用 `build_headers(meta)` 注入签名头。示例见 `adapters/` 下任一脚本的
+`build_headers` 实现；云端校验调用方 app 时需要按
+`meta.app_secret` 对时间戳+URL+请求体哈希做 HMAC 签名并附 `X-App-Sign` 等头。
+配好 `sources[].meta.{app_id, app_secret, app_agent}` 即可。
+

docs/configuration.md

@@ -0,0 +1,183 @@
+# 配置
+
+> 本文档由 README 拆分而来：网关配置、模型自动探测、多密钥。
+> 返回 [README](../README.md)。
+
+## 配置文件
+
+首次运行在 `-config` 路径自动生成默认配置（默认 `./config.yaml`），核心字段：
+
+```yaml
+listen: 127.0.0.1:8080        # 网关监听地址（建议绑内网/回环）
+gateway_keys: [sk-gw-<随机>]  # 初始 admin 密钥（每次安装随机生成，仅首启写入运行时存储用）
+default_model: AUTO           # model 无法路由时自动按优先级链选源
+adapter_dir: adapters         # Lua 适配目录；目录不存在时首启创建并 seed 内置，存在则只读
+runtime_file: runtime.json    # WebUI 编辑的源/密钥/AUTO 链持久化到此文件
+
+sources:
+  - name: deepseek
+    base_url: https://api.deepseek.com
+    api_key: sk-...            # 也可用 api_key_env: SOME_ENV 引用环境变量（不落盘明文）
+    adapter: deepseek
+    max_concurrent: 8
+    models:
+      - id: deepseek-v4-flash
+        priority: 100          # YAML 源首次启动会 seed 进 AUTO 链的初值
+        kind: chat
+      - id: deepseek-v4-pro
+        priority: 60
+        kind: chat
+    # 静态请求头（比适配器默认优先）
+    headers: { X-Tenant: prod }
+    # 传给 Lua build_headers 的透传元数据
+    meta: { app_id: x, app_secret: y }
+    temperature: 0.7
+    max_tokens: 4096
+    timeout: 120s              # 请求超时，默认 120s
+```
+
+
+## 自动探测模型列表
+
+不想手抄模型 id 时，只填 URL 和 key，让网关去扫：
+
+```bash
+# 只探测，不写任何东西
+curl -X POST http://127.0.0.1:8080/api/sources/discover \
+  -H "Authorization: Bearer sk-gw-…" -H "Content-Type: application/json" \
+  -d '{"base_url":"https://api.vendor.com/v1","api_key":"sk-…"}'
+# → {"ok":true,"base_url":"…","endpoint":"","models":["gpt-x",…],
+#    "image_models":["gpt-image-2"],"models_path":"/models","latency_ms":1387,…}
+
+# 探测后直接落库
+curl -X POST http://127.0.0.1:8080/api/sources/discover/apply \
+  -H "Authorization: Bearer sk-gw-…" -H "Content-Type: application/json" \
+  -d '{"base_url":"https://api.vendor.com/v1","api_key":"sk-…","name":"vendor"}'
+```
+
+WebUI 源编辑框里也有「自动探测模型」按钮：填好 URL 和 key 点一下，列表自动
+填好，仍需点保存才生效。
+
+几点值得知道：
+
+- **URL 形态不用猜。** 厂商对「`/v1` 放 base 还是放 path」没有共识，探测会
+  依次尝试若干候选并返回答对的那个，同时给出归一化后的 `base_url` +
+  `endpoint`（`endpoint` 平时留空，因为 openai adapter 自带 `/chat/completions`）。
+- **`image_endpoint` 是另一套规则。** 网关的图像默认路径是**绝对**的
+  `/v1/images/generations`，而 chat 默认是相对的。所以 `base_url` 带 `/v1` 时
+  必须显式给 `image_endpoint: /images/generations`，否则图像请求 404。
+  探测会替你填好。
+- **图像模型只看上游自己声明的 modality**，绝不从模型名猜。`gpt-image-2`
+  这类 id 也出现在同时提供对话服务的目录上，猜错会让 AUTO 链选中一个只支持
+  图像生成的模型，其后每一层都被它拖垮。
+- **不会逐模型探活。** 34 个模型的目录跑一轮要花钱且耗时数分钟，`/models`
+  是足够便宜的信号；模型是否真能应答留给正常请求去发现。
+- **对已存在的源，必须显式 `replace_models: true`** 才会覆盖模型列表。重新
+  探测顺手抹掉为 AUTO 链手工调过的 priority，是最难排查的一类数据丢失。
+- **`api_key` 留空表示沿用已存值**，不会把凭据清空。
+
+
+`config.yaml`、`runtime.json`、`master.key`、`adapters/` 均不提交版本库（见
+`.gitignore`），请随备份带走——`master.key` 丢失后运行时密文无法解密。
+
+运行文件生成在 `-config` 所在目录：`adapter_dir` / `runtime_file` 自动指向
+配置文件旁边的 `adapters/` 与 `runtime.json`，从任意工作目录启动都可用。
+
+运行时文件（`runtime_file`）中的敏感字段自动加密：
+
+- 加密算法 AES-256-GCM，格式 `enc:v1:<base64>`。
+- 主密钥来源：环境变量 `MODELROUTER_MASTER_KEY`（64 位 hex）；否则读取
+  `runtime_file` 同目录的 `master.key`；都不存在时首次启动自动生成 `master.key`（0600）。
+- 升级时旧明文文件自动兼容：首次运行正常读取，任何 UI 保存操作触发全文件加密迁移。
+- 注意：**master.key 丢失后密文无法解密**，请随配置一起备份；切勿提交到版本库。
+
+## 网关密钥（多密钥）
+
+网关鉴权采用「多密钥 + 角色 + 模型范围」架构，密钥持久化在 `runtime_file` 的
+`keys` 字段（加密存储）：
+
+- `gateway_keys` 配置只是**初始 admin 密钥种子**：首次启动迁移为运行时 admin
+  key，之后不再参与鉴权管理。
+- **重要：首次启动后请在 WebUI「密钥」页更换管理员密钥**——初始密钥明文写在
+  `config.yaml` 里（每次安装随机生成），继续使用存在被盗风险；用新密钥登录后删除初始密钥。
+- WebUI **密钥页**可创建/删除密钥；每个密钥可指定 `admin`（管理全部）或
+  `user`（仅看自己的 key）角色，并配置**模型范围**（模型 + 源 + token 配额 +
+  重置周期）。
+- 客户端用任意一个已授权的密钥明文作为 Bearer（`Authorization: Bearer <key>`）。
+- 删除密钥即从运行时存储移除，立即失效。
+
+### 密钥用量配额
+
+**配额按模型单独设置**：每个密钥的 `models[]` 里，每一条模型范围各自带一份
+token 配额与请求数配额。一个模型用满只影响该模型，同一密钥的其它模型照常工作。
+
+```yaml
+keys:
+  - key: sk-gw-<hex>
+    role: user
+    name: agent-alice
+    models:
+      - model: deepseek-v4-flash
+        token_quota: 1000000   # 本周期内该 key 用这个模型的 token 预算
+        req_quota: 20000       # 本周期内的请求次数
+        period: nhour          # "" | hour | week | month | nhour
+        hours: 6               # 仅 nhour
+      - model: AUTO
+        token_quota: 5000000   # AUTO 也是一条独立配额
+        period: day            # ← 这种写法会被拒绝（词表只有 hour/week/month/nhour）
+      - model: kimi-k3        # 未列配额 = 无限
+```
+
+- **没有「整钥总配额」**：这是刻意的设计。整钥总额会让「把 A 模型的额度挪给
+  B 模型」变成一次全局重分配；按模型独立计费则每个模型各自可控，运维可以
+  看出哪个模型吃掉了预算。
+- `period` 词表：空 = 永不过期（累计总量），`hour` / `week` / `month` = 固定窗口，
+  `nhour` + `hours` = 自定义小时数。**拼错的周期在写入时就被拒**，不会静默变成
+  永不过期。
+- 配额严格按密钥隔离，且**同一密钥内按模型隔离**：A 密钥用满 `m1` 不会消耗
+  B 密钥的额度，同一密钥的 `m2` 也不受影响。
+- 配额统计含聊天、流式、生图，跨重启从审计日志回放（保留 40 天，覆盖最长的
+  month 窗口）。
+- 配额耗尽返回 **429 + `Retry-After`**（`rate_limit_exceeded`），消息里点名是哪个
+  模型用满了，客户端可等窗口重置后自动恢复；模型越权才是 403。
+  **admin 密钥永不受配额限制**（其 scope 上的配额也不强制），避免把管理员
+  锁在门外。
+- 窗口用量按整点小时分桶统计，实际释放比配置窗口最多晚 1 小时（配额宁可晚释放
+  也不超发）。
+- `PUT /api/keys/{key}` 提交 `models` 即同时提交它们的配额（配额就是 scope 的一部分，
+  不存在会与模型列表脱节的第二份预算）。显式 `0` = 解除该模型的限制。
+
+#### 配额拒绝 vs 容量拒绝：两种「拒绝」含义不同
+
+配额在**入口**检查，不占用任何上游槽位，因此是廉价拒绝；容量不足则要走
+`busyWait` 有界等待后才返回 503。实测（容量 8、0.6s/请求、100 并发）：
+
+| 情况 | 状态码 | 延迟 | 是否打到上游 |
+|---|---|---|---|
+| 配额用尽（`req_quota` 50 → 100 并发） | `429 rate_limit_exceeded` + `Retry-After` | **19–21ms** | 否 |
+| 容量打满（同一拓扑，未限配额） | `503 upstream_error` | ~2.6s（`busyWait`） | 超出部分否 |
+
+客户端因此可以区分「等窗口重置」与「等上游腾出容量」：前者只需等待，后者
+通常该降并发或换源。
+
+#### 配额的开销
+
+配额桶按 (密钥, 模型, 整点小时) 分桶保留 40 天，实测（AMD 7840HS）：
+
+- 窗口查询按窗口长度而非保留总量扫描：1h 窗口 49ns、24h 窗口 55ns、
+  30d 窗口 3.9μs（此前全扫保留总量，960 桶时 5.9μs）。
+- 记录一条请求：283ns、3 分配（与引入配额前相同，分配来自 ring buffer）。
+- 内存：生产形态（7 密钥 × 8 模型 × 2 源 × 满 40 天 retention）约 **3.7MB**。
+  按源 pin 的 `source::model` 桶**惰性创建**——只有当某条配额真的 pin 了
+  某个源时才维护，否则每条记录多写一份桶，在 20 密钥 × 8 模型 × 3 源下会
+  多占 18MB。
+  代价：配置 pinned 配额之前的历史用量无法事后按源拆分，所以 **pinned 配额的
+  首个窗口可能少算**。
+
+```bash
+# 配额耗尽时客户端看到（点名了具体模型）
+HTTP/1.1 429 Too Many Requests
+Retry-After: 2100
+{"error":{"type":"rate_limit_exceeded","message":"token quota exceeded for \"deepseek-v4-flash\" (5000000/5000000)"}}
+```
+

docs/deployment.md

@@ -0,0 +1,142 @@
+# 部署与迁移
+
+> 本文档面向运维：单机 systemd 部署、发行包安装、目录与权限、从 `llmsproxy`
+> 改名后的迁移与回滚。
+> 返回 [README](../README.md)。
+
+## 1. 单机部署（推荐）
+
+仓库根目录的 `deploy.sh` 负责：构建 → 预检配置 → 备份 → 原子替换二进制 →
+重启服务 → 后置校验。它**不**修改源码，也**不**假设全新机器。
+
+```bash
+sudo ./deploy.sh                      # 默认读取仓库内 config.yaml
+sudo ./deploy.sh --config other.yaml  # 投放并部署另一份配置
+```
+
+部署目标（改名后）：
+
+> 两种布局并存，区别只在于二进制位置：**包安装**（deb/rpm）放 `/usr/bin`，
+> unit 自带；**源码部署**（`deploy.sh`）放 `/usr/local/bin`，需要让 unit 的
+> `ExecStart` 指向它。`deploy.sh` 会校验这一点，写错时直接中止而不是留下
+> 一个 203/EXEC 的服务。
+
+| 项 | 路径（源码部署 / 包安装） |
+|---|---|
+| 二进制 | `/usr/local/bin/modelrouter` / `/usr/bin/modelrouter` |
+| 配置 | `/etc/modelrouter/config.yaml` |
+| 适配器 | `/etc/modelrouter/adapters/` |
+| 插件（可选） | `/etc/modelrouter/plugins/`（需在配置里设 `plugin_dir`） |
+| 主密钥 | `/etc/modelrouter/master.key`（0600） |
+| systemd unit | `/etc/systemd/system/modelrouter.service` |
+
+`deploy.sh` 会在以下情况**中止**而不是"看起来成功"：
+
+- 部署后二进制起不来、或新配置 `-check` 不通过；
+- 凭据无法解封（`master.key` 丢失 / 与配置不匹配）；
+- 检测到改名前的 `/etc/llmsproxy`，但新配置尚未建立（否则会部署出一个
+  没有 key、没有源的空实例，所有请求都被拒）。
+
+## 2. 发行包安装
+
+三个产物由 `packaging/core-dist.sh` 生成到 `cmd/build/dist/`：
+
+```text
+modelrouter_<ver>_amd64.deb
+modelrouter-<ver>.x86_64.rpm
+modelrouter-<ver>-linux-amd64.tar.gz
+```
+
+deb / rpm 安装后：
+
+```bash
+sudo systemctl enable --now modelrouter
+sudo journalctl -u modelrouter -f
+```
+
+tar 包内只有：`modelrouter`、`config.example.yaml`、`modelrouter.service`、
+`migrate-to-modelrouter.sh`、`adapters/`。
+
+**发行包不含业务插件**（仓库只保留 `billing` / `usage` 两个与厂商无关的示例）。
+它们保留在仓库中作为可选扩展，需要时手工放入 `plugin_dir`，网关不会自动安装、
+覆盖或删除。这是有意的：插件属于运维资产，生命周期不与内核绑定。
+
+首次启动会在 `/etc/modelrouter/config.yaml` 生成一份**随机 admin key** 的配置，
+key 打印在启动日志里。发行包不预置任何可用 key（历史上曾因此泄露过一个真实的
+`sk-gw-local-0001`）。
+
+## 3. 目录与权限
+
+- `/etc/modelrouter` 是唯一的读写根（systemd unit 用 `ProtectSystem=strict` +
+  `ReadWritePaths=/etc/modelrouter` 固化这一点）。
+- `master.key` 必须 0600。当前 unit 以 root 运行：降权到普通用户需要先把
+  `master.key` 交给该用户并统一目录属主，否则服务会"带着明文落盘继续跑"，
+  失败只留一行日志——这是独立变更，不要和加固混在一起。
+- `plugin_dir` 由运维维护；插件自带的状态文件也在其中。
+
+## 4. 从 llmsproxy 迁移
+
+二进制、配置目录、服务名、环境变量已统一为 `modelrouter`：
+
+| 旧 | 新 |
+|---|---|
+| `/usr/local/bin/llmsproxy` | `/usr/local/bin/modelrouter` |
+| `/etc/llmsproxy/` | `/etc/modelrouter/` |
+| `llmsproxy.service` | `modelrouter.service` |
+| `LLMSPROXY_PLUGIN_DIR` | `MODELROUTER_PLUGIN_DIR` |
+| `LLMSPROXY_XDISPLAY_BASE` | `MODELROUTER_XDISPLAY_BASE` |
+| `LLMS_PROXY_MASTER_KEY` | `MODELROUTER_MASTER_KEY` |
+
+迁移脚本**默认 dry-run**，复制而非移动，旧目录永不删除：
+
+```bash
+sudo ./packaging/migrate-to-modelrouter.sh          # 预演（逐条打印将做什么）
+sudo ./packaging/migrate-to-modelrouter.sh --apply  # 执行
+```
+
+它做五件事：复制状态目录并改写配置里的旧绝对路径、复核 `master.key` 0600、
+安装新二进制（旧文件保留）、由旧 unit 生成新 unit（改写 ExecStart /
+WorkingDirectory / ReadWritePaths）、复制 `/usr/share` 资源；最后跑
+`modelrouter -check` 并 `systemctl daemon-reload`。
+
+确认新服务可用后，再手工清理旧布局（脚本不会删）：
+
+```bash
+sudo systemctl disable --now llmsproxy.service
+sudo rm -f /etc/systemd/system/llmsproxy.service
+sudo rm -rf /usr/local/bin/llmsproxy /etc/llmsproxy /usr/share/llmsproxy
+```
+
+回滚（不需要旧文件被删）：
+
+```bash
+sudo ./packaging/migrate-to-modelrouter.sh --rollback --apply
+```
+
+**读取侧兼容**：旧配置路径与旧环境变量仍被接受，所以"先换二进制、后改 unit"
+这种顺序不会立刻失效。但写入永远落在新路径，迁移完成后不要两台布局混用。
+
+## 5. 部署后验证清单
+
+```bash
+# 1) 进程与监听
+systemctl status modelrouter --no-pager
+ss -ltnp | grep modelrouter
+
+# 2) 鉴权生效（无 key 必须 401）
+curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8080/v1/models
+
+# 3) 带 key 能列出模型
+curl -s -H "Authorization: Bearer $KEY" http://127.0.0.1:8080/v1/models | head -c 400
+
+# 4) 真实请求走通（证明 LuaJIT 适配器路径未被打断）
+curl -s -H "Authorization: Bearer $KEY" \
+  -d '{"model":"AUTO","messages":[{"role":"user","content":"hi"}]}' \
+  http://127.0.0.1:8080/v1/chat/completions | head -c 400
+
+# 5) 仓库文件与生产文件一致（部署脚本的收尾检查）
+sha256sum /usr/local/bin/modelrouter build/modelrouter
+```
+
+第 4 步失败但第 3 步成功，通常是适配器或上游凭据问题，不是部署问题：
+查看 `journalctl -u modelrouter -n 100`。

docs/design-plugin-kernel-api.md

@@ -0,0 +1,117 @@
+# 插件可调用内核动作（R4-c）设计
+
+> 用户决策（2026-10-11）：**全开**（13 个内核动作全部可调）+ **设置页可配权限** +
+> **Go 侧 deadline 兜底**。
+
+## 1. 要解决的问题
+
+用户原话："现在这个接口不全，可以多给一些接口让 lua 调用内核。"
+
+现状（`__llmsproxy_adapters`）是我手写的**三个闭包**（install/update/remove），
+只能管适配器。内核还有 13 个公开动作（Reload / ResetHealth / AddSource /
+RemoveSource / SaveAutoRules / key 增删改 / 列表读取…），插件一个都调不到。
+
+而"自动孵化 / 自愈 / 联动"这类插件能力（例如：探测到上游全挂 → 触发重探 +
+清冷却；发现额度变化 → 重排 AUTO 链）恰恰需要它们。
+
+## 2. 关键认识：Go 调用可以超时，Lua 死循环不能
+
+这决定了整个设计，也是用户提"阻塞 + Go 兜底"的原因：
+
+| 形态 | 能否中断 | 处置方式 |
+|---|---|---|
+| Lua 里 `while true do end` | **不能**（LuaJIT 限制，已实测三次） | 只能停用插件 |
+| Lua 里 `__llmsproxy_kernel.call("reload")` | **能** | Go 侧 `select` + deadline，超时把错误返回给 Lua |
+
+因为 Lua 只是在等 Go 的返回值，**不占用 Lua 栈的执行**。所以给每次内核调用加
+Go 侧 deadline 是可行的、干净的 —— 不需要"Lua 投递请求、Go 队列执行"：那样
+还要定义"超时后队列里积压的命令怎么办"，而 deadline 直接返回 `(nil, err)`，
+语义闭合。
+
+## 3. 依赖方向
+
+`internal/lua` **不能** import `internal/core`（方向是 core → lua）。
+按既有先例（`SetOutboundHosts`）注入一个**接口**，由 core 实现：
+
+```go
+// internal/lua/kernel_api.go
+type KernelAction struct {
+    Name        string   // "reload" / "reset_health" / ...
+    Title       string   // 设置页展示用
+    Description string
+    Danger      string   // ""=只读, "write"=改配置, "danger"=破坏性
+    Params      []string // 参数名（设置页/文档用）
+}
+
+type KernelAPI interface {
+    // Actions 列出可调用动作（设置页据此渲染权限开关）
+    Actions() []KernelAction
+    // Invoke 执行一个动作，args 为已解码参数。**必须**尊重 ctx 的 deadline。
+    Invoke(ctx context.Context, name string, args map[string]any) (any, error)
+}
+```
+
+core 侧实现 `Invoke` 时用**现有的 Core 方法**，本身不持 `p.mu`，因此不会与
+插件锁成环（已核实：所有 `ps.mu` → `p.mu` 都是先释放 `ps.mu`）。
+
+## 4. 权限模型
+
+复用 `visibilityOverrides` 的**同一套持久化机制**（`.access.json`），
+但用独立的键空间：`kernel:<plugin>` → 逗号分隔的动作名或 `*`。
+
+- **默认拒绝**：插件不显式获得授权时，`call` 返回 `(nil, "not permitted")`。
+  理由：全开是"能力存在"，不是"默认授予"。一个插件能删密钥这件事必须由
+  运维在设置页点过确认。
+- **`danger` 类必须显式逐个授权**，不接受 `*`。这一条是防止运维图省事写
+  通配符，把"删密钥"和"读状态"等同起来。
+- 插件可以在自己源码里**声明**它想要的动作（`plugin.kernel = {...}`），
+  设置页把它显示为"申请"，运维批准后才生效。声明不等于授权。
+
+## 5. 动作清单（13 个 + 适配器 3 个）
+
+| 动作 | 危险级 | 对应 Core 方法 | 超时 |
+|---|---|---|---|
+| `status` | 只读 | 读 AutoSlotStates/Registry | 2s |
+| `list_sources` | 只读 | Sources() | 2s |
+| `list_models` | 只读 | Registry().ModelList() | 2s |
+| `list_adapters` | 只读 | ListAdapters() | 2s |
+| `adapter_owner` | 只读 | AdapterOwnerOf | 2s |
+| `reload` | write | Reload() | 30s |
+| `reset_health` | write | ResetHealth() | 5s |
+| `probe_all` | write | Registry().ProbeAll(ctx) | 60s |
+| `save_auto_rules` | write | SaveAutoRules() | 10s |
+| `save_key_auto` | write | SaveKeyAuto() | 10s |
+| `add_source` | write | AddSource() | 30s |
+| `remove_source` | danger | RemoveSource() | 30s |
+| `create_key` | write | CreateKey() | 10s |
+| `update_key` | write | UpdateKey() | 10s |
+| `delete_key` | danger | DeleteKey() | 10s |
+| `install_adapter` | write | UploadAdapter() | 10s |
+| `remove_adapter` | danger | RemoveAdapter() | 30s |
+| `set_plugin_enabled` | danger | SetEnabled() | 5s |
+
+**绝不开放**：`ListKeys()` / `FindKey()` —— 返回**明文密钥**。给插件读密钥等于
+把网关的凭据面交给第三方 Lua。要用 key 相关动作就只用带掩码的形态。
+
+## 6. Lua 调用面
+
+```lua
+-- 探测：拿到可用动作与自己的授权情况
+local acts = __llmsproxy_kernel.actions()
+-- 调用：统一入口，返回 (result, err)
+local ok, err = __llmsproxy_kernel.call("reset_health", {})
+local st, err = __llmsproxy_kernel.call("status", {})
+```
+
+统一入口（而不是每个动作一个全局函数）的理由：动作集合将来会变，而
+`__llmsproxy_kernel` 的签名不变 —— 否则每加一个动作就要动
+`bootPluginState`，而那里正是历史上出过 cgo 崩溃的地方。
+
+## 7. 实施顺序
+
+1. `internal/lua/kernel_api.go`：接口 + 动作元数据 + 权限检查 + deadline 包装
+2. `internal/lua/plugin_kernel_bridge.go`：Lua 绑定（`actions` / `call`）
+3. `internal/core/kernel_actions.go`：实现 `KernelAPI`（core 侧）
+4. 权限持久化：复用 `.access.json`，新键空间
+5. 设置页：插件行加"内核动作"列 + 授权编辑
+6. 测试：Go 侧全动作、Lua 侧调用、权限拒绝、deadline 生效、危险动作需显式授权

docs/development.md

@@ -0,0 +1,34 @@
+# 开发与目录结构
+
+> 本文档由 README 拆分而来：目录职责、开发协作、测试。
+> 返回 [README](../README.md)。
+
+## 目录
+
+```
+cmd/modelrouter            # 入口（纯 Go 服务端）
+cmd/gui                  # Electron 桌面 GUI（内嵌核心 + 托盘 + 打包）
+internal/config          # YAML 配置加载/校验
+internal/lua             # LuaJIT VM + worker 池 + AdapterCache + 内置适配器 (embed)
+internal/provider        # Provider(HTTP) + Registry(路由)
+internal/gateway         # OpenAI 兼容 HTTP 服务 + 鉴权 + SDK/流式 + WebUI
+internal/types           # 统一格式 & OpenAI wire 类型
+```
+
+## 开发协作
+
+- **Git 分支工作流**：主分支 / 特性分支 / 发布分支的规范见
+  **[docs/git-workflow.md](docs/git-workflow.md)**（[English](docs/git-workflow-en.md)）。
+  主线：`main` 永可部署 → 特性合入 `main` → 切 `release/vX.Y.Z` 打 tag 发布 →
+  hotfix 提交发布分支并 cherry-pick 回 `main`。
+- **Lua 适配器协议**：见 **[docs/lua-adapters.md](docs/lua-adapters.md)**
+  （[English](docs/lua-adapters-en.md)）。
+
+## 测试
+
+```bash
+go test -tags luajit ./...
+```
+
+覆盖：配置校验、适配器加载/变换、签名钩子、Gateway 鉴权、SDK 结算、SSE 流式、
+模型路由、多模态透传与 thinking 预算。

docs/features.md

@@ -0,0 +1,48 @@
+# 实现与完整特性
+
+> 本文档由 README 拆分而来：架构来源与完整特性清单（README 只保留摘要）。
+> 返回 [README](../README.md)。
+
+## 实现
+
+统一 OpenAI 兼容网关：把多个上游 LLM 源（DeepSeek、Qijiar、OpenAI、Anthropic、Gemini、
+Groq、Mistral、Ollama、KimiCode…）通过 **Lua 适配器** 做协议转换，对内网暴露
+一个标准 `OpenAI Chat Completions` 接口（`/v1/chat/completions` + `/v1/models`），
+支持单次结算、SSE 流式、**AUTO 模型路由**与**多模态**透传。
+
+从上游多源 LLM 适配层
+（`internal/agent/api/provider.go` + `internal/lua/adapters/*`）抽离而来并独立演进。
+
+## 特性
+
+- **多源**：一个进程内配置任意多个上游源，按请求的 `model` 自动路由。
+- **AUTO 模式**：`default_model: AUTO` 时按优先级页保存的 AUTO 链档位逐档调度并发请求。
+- **统一输出**：所有源对外都是 OpenAI 格式（含 `reasoning_content`、`tool_calls`、`usage`）。
+- **生图**：`POST /v1/images/generations`，`kind: image` 的模型独立路由。
+- **多模态**：`content` 数组（`image_url` 等）在多源间无损透传，Anthropic/Gemini/Ollama 自动转换。
+- **LuaJIT VM**：基于 golua 绑定的 LuaJIT，每个适配器独立 VM + worker 池，安全并发。
+  池是 **弹性** 的：`Σ max_concurrent` 只是上限而非预分配，状态按需创建、空闲时自动回收，
+  因此空载网关几乎不持有 Lua 状态（启动 0 个）。
+  - 扩容步长由适配器最大并发决定：`clamp(ceil(max/8), 1, 8)`，且仅在**真实并发争用**
+    （所有现有状态均已占用）时批量预热，顺序流量始终只用 1 个状态。
+  - 缩容步长由当前连接数决定：`clamp(ceil(冗余/(1+占用)), 1, 冗余)`——无连接时一轮收到
+    常驻下限，忙时每轮只释放 1 个，保护热路径。WebUI「适配器」页展示实时池指标。
+- **thinking 预算**：请求里的推理开关/预算归一为中立的 `thinking:{enabled,budget_tokens}`，
+  适配器映射到各自协议（Anthropic `thinking.budget_tokens`、Gemini `thinkingConfig`、
+  DeepSeek `thinking.type="disabled"`）。
+- **Lua 适配协议**：每个源挂一个 `.lua` 适配器，完成 `transform_request` /
+  `transform_response` / `transform_stream_chunk` 双向转换，协议差异全在 Lua 层。
+- **签名 / 请求头钩子**：适配器可定义 `build_headers(meta)`，在 Go 发 HTTP 前
+  动态注入/签名请求头——用于云端 API 校验调用方 app（如 KimiCode 只放行特定
+  agent）。提供 `hmac_sha256_hex` / `sha256_hex` / `base64_encode` 等签名辅助。
+- **WebUI**：内置管理界面，可在线查看/新增/编辑上游源与模型、配置 AUTO 优先级链、
+  管理密钥模型范围，写入运行时文件持久化。
+- **密钥加密存储**：运行时文件中的上游 `api_key`、自定义请求头值、网关 key 均以
+  AES-256-GCM 加密落盘（`master.key` 独立 0600，或 `MODELROUTER_MASTER_KEY`）。
+- **实时源探测**：状态页探测各源可达性（`GET {base}/models`，失败回退最小请求），
+  不污染正常调度的退避状态，错误信息在 UI 可悬停查看。
+- **鉴权**：网关自身用 `gateway_keys` 校验客户端 Bearer key；与上游各自的 key 相互独立。
+- **流式**：SSE `chat.completion.chunk`，含角色首包与 `[DONE]` 收尾。
+- **全链路审计**：HTTP 访问日志、登录/配置变更审计、请求记录 CSV 导出、密钥用量统计 CSV 导出
+- **AES-256-GCM 加密存储**：运行时文件敏感字段加密落盘，master.key 0600 权限，支持环境变量注入主密钥
+

docs/git-workflow-en.md

@@ -134,7 +134,7 @@ git push origin vX.Y.Z
 make core-dist VERSION=X.Y.Z
 
 # 5. create GitCode Release + upload
-#    https://gitcode.com/JianFeeeee/ModelRouter/releases
+#    https://atomgit.com/HeNanMU-EEA/ModelRouter/releases
 #    upload MUST use PUT --http1.1, otherwise the OBS callback fails and the
 #    file never lands in the release; replacing artifacts = delete the tag
 #    (release disappears with it) then recreate.

docs/git-workflow.md

@@ -124,7 +124,7 @@ git push origin vX.Y.Z
 make core-dist VERSION=X.Y.Z
 
 # 5. 创建 GitCode Release + 上传安装包
-#   见 https://gitcode.com/JianFeeeee/ModelRouter/releases
+#   见 https://atomgit.com/HeNanMU-EEA/ModelRouter/releases
 #   上传注意：必须 PUT --http1.1，否则 OBS 回调失败文件不进列表；
 #   替换旧版附件 = 删 tag（release 随 tag 一起消失）后重建。
 ```

docs/gui.md

@@ -0,0 +1,86 @@
+# 桌面 GUI（Electron）
+
+> 本文档由 README 拆分而来：Headless 与 Desktop 两种产物、构建与运行。
+> 返回 [README](../README.md)。
+
+## 桌面 GUI（Electron）
+
+可选独立产物，面向内网桌面端用户：**内嵌 ModelRouter 核心**（桌面壳形态），
+一键启停、系统托盘常驻、开机自启、静默启动，内容区一比一内嵌完整 WebUI
+（状态/对话/密钥/优先级/源/适配器 六个页面，免登录）。服务器用户继续使用纯 Go 二进制。
+
+### 两种安装包：Headless 与 Desktop
+
+发布到 Release 的安装包分为两类，面向不同用户：
+
+- **Headless（服务器版）**：纯核心二进制（无 Electron），配置驱动，适合服务器、
+  容器、systemd 等无人值守场景。发布包含 `deb` + `rpm`（含 systemd unit、内存调优
+  与示例配置）与 `tar.gz`（二进制 + `config.example.yaml` + `adapters/` + README），
+  Windows 为 `.zip`。
+- **Desktop（桌面版）**：Electron GUI 安装包，内嵌核心，适合个人桌面日常使用。
+
+### 选择 GUI 还是纯后端
+
+| 场景 | 推荐 | 理由 |
+| ------ | ------ | ------ |
+| 服务器 / 内网网关 / 无人值守常驻 | **Headless**（单二进制） | 极轻量（二进制 ~8MB；RSS 随源数量，单源 ~10MB、16 源 ~32–35MB，见[内存占用](#内存占用实测与调优)），零依赖单进程，直接跑在 systemd/任意容器里，远程管理 |
+| 个人桌面日常使用 / 多设备内网共享 | **Desktop**（GUI） | 免登录内嵌 WebUI、系统托盘一键启停、开机自启、静默后台，适合不懂命令行的使用者 |
+| Windows 桌面 | **Desktop** | 纯后端在 Windows 上需自行注册服务，GUI 提供原生托盘/自启体验 |
+| CI 一键出三平台安装包 | **Desktop 打包脚本** | `make gui-*` 系列出 deb/AppImage/NSIS，可进发行流水线 |
+
+两者完全同源：GUI 内嵌的就是纯后端的同一份 `modelrouter` 核心（LuaJIT 版），
+配置/适配器格式完全一致，可随时互换。
+
+### 构建与运行
+
+```bash
+cd cmd/gui
+npm install
+make gui          # 或 npm run dev —— 开发运行（需先 make build 生成 bin/modelrouter）
+make gui-deb      # 本机构建 deb（分享给其它 Linux 用户）
+make gui-dist     # 构建 deb + rpm + AppImage（linux，rpm 需系统 rpmbuild）
+make core-dist    # 构建核心 deb + rpm + tar.gz（服务器版，需 nfpm）
+make gui-win      # 构建 Windows nsis 安装包（本机 mingw + wine）
+make gui-win-docker  # 构建 Windows nsis 安装包（全 docker 自足，含 wine）
+make gui-dist-dir # 仅解包产物（调试用，不产安装包）
+```
+
+产物在 `cmd/build/gui-dist/`：`ModelRouter-<ver>.AppImage`、
+`modelrouter-gui_<ver>_amd64.deb`、`ModelRouter Setup <ver>.exe`（win）；
+核心包在 `cmd/build/dist/`：`modelrouter_<ver>_amd64.deb`、`modelrouter-<ver>.x86_64.rpm`。
+rpm 需系统 `rpmbuild`。每次打包后都会跑 `packaging/verify-dist.sh` 产物体检
+（大小下限门禁——1.3.0 曾因宿主 wine 损坏产出 264KB 的残缺 exe 却没有被拦住）。
+发布时会把产物重命名为 `ModelRouter-Desktop-*`（GUI）与 `ModelRouter-Headless-*`（纯后端）
+两类上传到 GitCode Release。
+
+#### Windows 打包（docker 固化方案，推荐）
+
+`make gui-win-docker` 一步完成：交叉编译核心 + NSIS 打包**全在 docker 内**，宿主机
+**无需 mingw、无需 wine、无需 node**。镜像内置 wine（含 i386 运行时），修复了
+此前宿主 wine 损坏（缺 syswow64 → `c0000135`）导致产出 264KB 残缺 exe 的问题。
+
+- 构建镜像：`cmd/gui/docker/win-builder/Dockerfile`（`golang:1.25` + mingw-w64 +
+  node + wine32/wine64 + 预编译 LuaJIT windows 版 `lua51.dll` 与 import lib）。
+  LuaJIT 源码取自国内镜像，上游兜底；Go 依赖走 `https://goproxy.cn`。
+- 一键发布：`cmd/gui/scripts/dist-win-docker.sh` —— 核心编译 + electron-builder
+  NSIS 全部在容器内执行（`USE_SYSTEM_WINE=true` 走镜像内的 wine）；electron-builder
+  的 ~400MB 工具链缓存挂到 docker volume，二次构建增量复用。
+- 首次运行自动构建 `modelrouter/win-builder` 镜像（约 5–10 分钟）；已存在则直接复用。
+- 打包完成后自动跑 `packaging/verify-dist.sh` 门禁：exe 低于 5MB / deb·rpm·7z 低于
+  10MB 即失败——残缺安装包不可能再溜进 Release。
+
+```bash
+# 全 Windows 发布（核心重编 + NSIS）：
+make gui-win-docker
+```
+
+### 特性
+
+- **内嵌核心**：自动拉起捆绑的 `modelrouter`（LuaJIT 版），配置/密钥/adapter 存于
+  用户数据目录 `<userData>/profile/`，首次运行自动生成随机 admin key 写入 `keys`
+  （非 seed）并自注入，WebUI 免登录、不会提示"更换初始密钥"。端口可在设置中更改（默认 8787）。
+- **系统托盘**：内核状态点、开机自启/静默启动开关、启停/重启内核、退出。
+- **静默启动**：`--silent` 参数或设置项，启动仅驻托盘不弹窗；首次运行始终显示窗口。
+- **开机自启**：Windows/macOS 用 `setLoginItemSettings`；Linux 写
+  `~/.config/autostart/modelrouter-gui.desktop`，静默开启时自动附加 `--silent`。
+

docs/lua-adapters.md

@@ -23,7 +23,7 @@
 5. [内置辅助函数](#5-内置辅助函数)
 6. [meta 约定](#6-meta-约定)
 7. [一个完整的最小适配器](#7-一个完整的最小适配器)
-8. [多模态与 disable_thinking](#8-多模态与-disable_thinking)
+8. [多模态与 thinking 预算](#8-多模态与-thinking-预算)
 
 ---
 
@@ -73,8 +73,9 @@ end
 
 约定：
 - 返回的字符串将被原样 POST 到 `base_url + (source.endpoint or adapter.endpoint)`。
-- 请求体中包含的 `disable_thinking`、`extra_body` 等网关内部字段，转换时应自行清理
+- 契约 v2 起，内核**不再向适配器泄漏任何私有字段**：请求体里不会出现
-  （`openai.lua` 即删掉这两个字段）。
+  `disable_thinking` / `extra_body` 这类键。推理开关与预算用中立的 `thinking`
+  字段表达（见下），适配器按目标协议各自映射。
 
 ### `transform_response(raw_body) -> string`
 
@@ -176,6 +177,81 @@ end
 
 返回非 table 会报错；返回 `{}` 表示无自定义头（不会回退静态头）。
 
+### 出向序列化钩子（契约 v2）
+
+默认情况下，网关把适配器产出的**统一格式**按 OpenAI 线格式序列化给下游（这是内核
+保留的兑底，R5）。当上游协议不是 OpenAI（例如 Anthropic Messages API）时，适配器
+可以接管出向序列化，自己决定写给客户端的字节。
+
+五个钩子都是可选的，**全不声明就沿用 OpenAI 兑底**。一旦声明某个钩子，它就完全
+接管那一段输出——内核不会再往里混合 OpenAI 字段（不是“叠加”，是“取代”）。
+
+| 钩子 | 何时被调用 | 返回值 |
+|------|-----------|--------|
+| `serialize_response(payload)` | 非流式响应 | 写给客户端的响应体字符串 |
+| `serialize_open_preamble(payload)` | 流式开始（发第一个 chunk 前） | 要发的字节（可为 `""`） |
+| `serialize_chunk(delta)` | 流式每个增量 | 要发的字节（`""` 表示跳过） |
+| `serialize_close(payload)` | 流式结束 | 收尾字节（如 SSE `message_stop`） |
+| `serialize_done_marker(payload)` | 流式结束处的 `[DONE]` 标记 | 要写的字节；不声明则用 OpenAI 的 `data: [DONE]` |
+
+`payload` 是统一格式的结果表（见 `internal/types/maxwire.go` 的 `UnifiedResult` /
+`UnifiedDelta`），字段名中立，不依赖具体上游协议。
+
+```lua
+-- anthropic.lua 的简化形式：把统一结果写成 Anthropic Messages 形状
+function adapter.serialize_response(payload)
+  return json.encode({
+    id = payload.id,
+    type = "message",
+    role = "assistant",
+    content = { { type = "text", text = payload.content or "" } },
+    stop_reason = payload.finish_reason == "stop" and "end_turn" or nil,
+    usage = { input_tokens = payload.prompt_tokens, output_tokens = payload.completion_tokens },
+  })
+end
+```
+
+注意：钩子返回的是**完整要发送的字节**（包括必要时的 `event:` / `data:` 前缀与
+空行），因为只有适配器知道目标协议的 SSE 细节。
+
+### 入向让出钩子（契约 v2，`serialize_*` 的对偶）
+
+出向是「统一格式 → 客户端协议」，入向是它的镜像：「客户端协议 → 统一格式」。
+内核为对外端点（如 `POST /v1/messages`）**保留一份内置解析作为兜底（R5）**，
+但适配器可以**声明自己服务该端点**并接管解析——那时内核主动让出，客户端协议的
+迭代就完全落在适配器里，不再需要动内核。
+
+声明分两步：
+
+1. `adapter.ingress = "/v1/messages"` —— 声明**认领**哪条对外入站路径。
+   同一条路径被多个适配器认领是配置错误；内核按适配器名字典序确定性选择。
+2. `deserialize_request(raw_body) -> unified_json` —— 把客户端原始 body 解析成
+   统一请求（`UnifiedRequest` 的 JSON），后续调度与 `/v1/chat/completions` 完全一致。
+
+规则：
+
+- 只有声明了该端点且实现了 `deserialize_request` 的适配器会让内核让出。
+- 没有适配器认领时，内核内置解析接管（未配置的部署仍能工作）。
+- 认领了但钩子**执行失败**是硬错误（400），不会静默回落到另一种协议解析——
+  客户端说的正是这个适配器声称支持的协议，悄悄换解析器只会错误调度请求。
+
+```lua
+-- anthropic.lua 的简化形式：把 Anthropic Messages 请求解析成统一格式
+adapter.ingress = "/v1/messages"
+function adapter.deserialize_request(raw_body)
+  local req = json.decode(raw_body)
+  local msgs = {}
+  for _, m in ipairs(req.messages or {}) do
+    table.insert(msgs, { role = m.role, content = m.content })
+  end
+  return json.encode({
+    model = req.model, messages = msgs, max_tokens = req.max_tokens,
+    system = type(req.system) == "string" and req.system or "",
+    stream = req.stream == true,
+  })
+end
+```
+
 ## 5. 内置辅助函数
 
 所有适配器共享的全局函数（由 Go 注入）：
@@ -219,8 +295,13 @@ adapter.headers = { ["X-Tenant"] = "prod" }
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
-    req.disable_thinking = nil
+    -- 契约 v2：内核不再向适配器泄漏私有字段，`thinking` 是中立语义字段。
-    req.extra_body = nil
+    -- 这里把中立的 thinking 映射成目标协议需要的形状（示例：OpenAI 家族用
+    -- reasoning_effort）。不关心推理的源可以直接忽略它。
+    if req.thinking ~= nil then
+        if req.thinking.enabled == false then req.reasoning_effort = "none" end
+        req.thinking = nil
+    end
     return json.encode(req)
 end
 
@@ -268,11 +349,21 @@ sources:
         kind: chat
 ```
 
-## 8. 多模态与 disable_thinking
+## 8. 多模态与 thinking 预算
 
 - **多模态**：请求的 `content` 若是数组（`[{type:"text"...},{type:"image_url"...}]`），Go 层
   原样透传给适配器。协议不支持的源（Anthropic/Gemini/Ollama）需在
   `transform_request` 内转换；内置 `anthropic.lua`/`gemini.lua`/`ollama.lua` 已实现。
-- **disable_thinking**：网关将请求体中的 `disable_thinking:true` 透传进
+- **thinking**：客户端请求里的推理开关/预算被内核归一成中立的 `thinking` 字段：
-  `transform_request`。DeepSeek 适配器将其映射为 `extra_body.thinking = { type = "disabled" }`
+
-  并清理自身字段；其余适配器按各自协议处理（可自行实现或忽略）。
+  ```json
+  {"thinking": {"enabled": true, "budget_tokens": 4096}}
+  ```
+
+  `enabled=false` 表示关闭推理，`budget_tokens` 可选。适配器自行映射到目标协议
+  （内置实现：Anthropic → `thinking.budget_tokens`、Gemini →
+  `generationConfig.thinkingConfig.thinkingBudget`、DeepSeek →
+  `thinking.type = "disabled"`）。不支持的源直接忽略该字段。
+
+  契约 v1 的内核私有字段 `disable_thinking` 已被取代——它曾是内核细节泄漏进
+  适配器请求体的典型例子（迫使 10 个适配器各写一行 `req.disable_thinking = nil`）。

docs/performance.md

@@ -0,0 +1,60 @@
+# 性能与内存
+
+> 本文档由 README 拆分而来：实测内存占用、调优手段、thinking 预算。
+> 返回 [README](../README.md)。
+
+## 内存占用（实测与调优）
+
+内存不是一个定数：它随**配置的源数量**增长（每个源一个 `http.Transport` 与连接池、
+一组健康状态），与运行时长无关。本机实测（Linux x86_64，12 核）：
+
+| 部署形态 | 启动 RSS | 稳态 RSS |
+|---|---|---|
+| 1 源 / 1 适配器（最小配置） | ~4 MB | ~10 MB |
+| 1 源 + 29 MB 历史审计日志 | ~19 MB | ~20 MB |
+| **16 源 / 13 适配器 / 59 模型（本机生产）** | ~28 MB | **~32–35 MB** |
+
+> 历史参考：本项优化前同一生产配置为 **~105 MB**。降幅来自三处：审计日志的**原始记录**
+> 不再常驻内存（约 25 MB；统计聚合仍扫全量，但扫完即释放）、Lua 状态池不再单调增长、
+> 以及下面两个运行时开关。
+
+内存构成（生产实例分段测量，`/proc/<pid>/smaps`）：
+
+| 区域 | RSS | 说明 |
+|---|---|---|
+| Go 堆 | ~14 MB | provider/registry/scheduler 结构 + 连接池缓冲 |
+| 其他匹名（线程栈 / LuaJIT chunk / runtime） | ~12 MB | 与线程数、已加载适配器数相关 |
+| 二进制 text+rodata | ~8 MB | 映射的可执行文件页（只读、可被内核回收） |
+| Go runtime 预留 | ~4 MB | VSZ 上看到的 2 GB+ 是地址空间预留，不占物理内存 |
+| 共享库 | ~3 MB | libc / libluajit / libm |
+
+**两个推荐的部署开关**（只适用于环境变量，无需改代码）：
+
+```ini
+# /etc/systemd/system/modelrouter.service
+Environment=GOGC=50
+Environment=MALLOC_ARENA_MAX=2
+```
+
+- `MALLOC_ARENA_MAX=2`：LuaJIT 的分配走 cgo → glibc malloc，glibc 默认允许 `8×nproc`
+  个 per-thread arena，每个碰到 malloc 的 OS 线程会占用一个（各约 1 MB，且**不归还给系统**）。
+  实测从 8–15 个 arena（约 7–12 MB）降到 **0**。
+- `GOGC=50`：把 Go 堆增长目标减半。**单独使用无效**（省下的堆会立即被更多 glibc
+  arena 吃掉，实测 20.3 → 21.5 MB 反而变大），必须与 `MALLOC_ARENA_MAX` 配合，
+  两者同时开启才降 **~19%**。网关是 I/O 密集型（本机 9 小时仅消耗 1min10s CPU），
+  多出的 GC 周期与它的空闲 CPU 相比可忽略。
+
+> 桌面版（Electron）已在 `cmd/gui/main.js` 里默认为内嵌核心注入这两个开关；手动 export 同名
+> 环境变量可覆盖。如果你用自己的 systemd unit / 容器，建议照上面加上。
+
+可选：`GOMEMLIMIT=48MiB` 作软上限，实测再省 ~1 MB；代价是逐近上限时 GC 转激进，
+温和场景不必开。
+
+## thinking 预算
+
+请求体带 `"thinking": {"enabled": false}` 时关闭推理；`"thinking": {"enabled": true,
+"budget_tokens": 4096}` 指定预算。内核只定义这个**中立**语义字段，映射到具体协议由
+适配器负责（Anthropic `thinking.budget_tokens`、Gemini `thinkingConfig.thinkingBudget`、
+DeepSeek `thinking.type = "disabled"`）。内核不再向适配器泄漏 `disable_thinking` /
+`extra_body` 这类私有字段。
+

docs/plugins.md

@@ -15,6 +15,37 @@ ModelRouter 的插件是**单个 `.lua` 文件**，放在 `config.yaml` 的 `plu
 
 ---
 
+## 0. 插件 SDK：唯一的公开接口
+
+插件与网关之间**只有一个**公开接口表 `modelrouter`。它不是 Go 包的导出符号，
+而是内核在启动插件 VM 时注入的 Lua 表；插件通过它调用内核能力，永远不需要
+（也不应该）依赖任何 `__llmsproxy_*` 全局、Go 结构体或内部函数名。
+
+```lua
+modelrouter.sdk_version   -- 字符串，契约版本；SDK 形状变化才递增
+modelrouter.http(spec)    -- 受出站白名单约束的 HTTP 请求
+modelrouter.secrets       -- 本插件专用的凭据保险库（set/get/delete/keys）
+modelrouter.kernel        -- 内核动作面（actions/call，受运维授权约束）
+modelrouter.adapters      -- 安装/更新/删除本插件自己的适配器
+modelrouter.plugin_dir    -- 本插件所在目录（内核给出的 ground truth）
+```
+
+规则：
+
+- **只依赖上表**。内核重命名内部全局、改实现或加守卫时，插件不受影响；
+  这正是把六个散落的全局收进一张表的原因。
+- **能力可能缺失**。未授予的能力（例如出站 HTTP 全禁）在表里是 `nil` 而不是
+  "存在但一调用就报错"，插件应先判断再使用。
+- **`sdk_version` 用来做兼容分支**。增量字段不递增版本；字段删除或调用约定
+  变化才递增。
+- **历史别名**：`llmsproxy`（旧产品名）与 `__llmsproxy_{http,secrets,kernel,
+  adapters,plugin_dir}` 仍然指向同一批对象，只为已安装的插件继续可用而保留，
+  已废弃；新插件使用 `modelrouter.*`。
+- 环境变量：插件目录优先读 `MODELROUTER_PLUGIN_DIR`，回退旧名
+  `MODELROUTER_PLUGIN_DIR`。
+
+---
+
 ## 1. 快速上手
 
 一个最小的插件：
@@ -46,7 +77,7 @@ plugin.ui = {
   page = {
     page_id = "hello",     -- kebab-case
     title = "Hello",
-    icon = "👋",
+    icon = "",
     order = 90,            -- 侧栏排序
     mount = [[<div id="hello">hello</div>]],
   },
@@ -65,16 +96,14 @@ return plugin   -- 必须返回一个 table
 core.New
   └─ lua.NewVM(adapter_dir).Start()        适配器状态
   └─ lua.NewPlugins(vm, plugin_dir)
-       ├─ SeedBundled()                    仅当目录不存在时写入内置插件（目前是 billing）
+`plugin_dir` 中的 Lua 文件由网关逐个加载；安装包不携带业务插件，也不会自动创建或写入插件文件。首次启用时，先把需要的 `.lua` 插件放入该目录，再启动网关。
-       └─ LoadDir()                        按文件名字典序逐个加载
 ```
 
 **加载失败不影响网关启动。** 一个语法错误的插件会被记录在 `GET /api/plugins`
 的 `error` 字段里，永远不会被调用。这与适配器一致，但理由更强：插件是可选的
 第三方扩展，因为一个 `.lua` 打错字就让网关起不来是错误的取舍。
 
-**目录一旦存在就是权威的。** 与适配器同规则：首启会 seed 内置插件，之后目录里
+插件目录由运维独立管理。网关不会 seed、覆盖或删除其中的插件；目录缺失时只会跳过插件加载，不影响网关启动。
-的文件说了算，删除或编辑内置插件都是真实生效的操作。
 
 ### 2.1 热更新
 
@@ -211,7 +240,7 @@ plugin.ui = {
   page = {
     page_id = "billing",   -- 必填,kebab-case。侧栏 data-tab 与 #tab-billing
     title = "Billing",     -- 必填,侧栏文字
-    icon = "💰",           -- 可选
+    icon = "",           -- 可选
     order = 40,            -- 侧栏排序,默认 100
     mount = [[...HTML...]],
   },
@@ -294,27 +323,191 @@ PUT    /api/plugins/{name}/state     替换状态（admin）
     "loaded": true
   }],
   "hook_errors": { "request_end": { "count": 3, "last_error": "billing: ..." } },
-  "plugin_dir": "/etc/llmsproxy/plugins"
+  "plugin_dir": "/etc/modelrouter/plugins"
 }
 ```
 
 `hook_errors` 是**排错入口**：一个坏掉的插件表现为"功能缺失"而不是报错，
 这个计数让它可见。
 
+### 5.4 定时任务（`PLUGIN.schedule`）—— 浏览器关着也会跑
+
+要求插件做“随时间发生”的事（定时刷新、token 轮换、指标采集）时，声明一个
+调度表：
+
+```lua
+plugin.schedule = {
+  { fn = "refresh_token", every = 3600 },                 -- 每小时（秒）
+  { fn = "daily_refresh", every = 86400, at = "03:30" }, -- 每天固定时刻（本地时间）
+  { fn = "poll_quota",    every = 300, jitter = 30 },     -- 首次触发加 0-30s 抖动
+}
+function plugin.refresh_token()
+  -- 注意：直接写 plugin.state 就是对的 ——
+  -- 任务和钩子跑在同一个 VM 上
+  plugin.state.token = fetch_new_token()
+end
+```
+
+| 字段 | 含义 |
+|---|---|
+| `fn` | 要调用的插件函数名（必须存在） |
+| `every` | 间隔秒数。与 `at` 二选一 |
+| `at` | 每天本地时间 `"HH:MM"`；设了它就忽略 `every` |
+| `jitter` | 首次触发额外随机延迟秒数（避免多实例整点齐射） |
+| `timeout` | 解析保留但目前**不生效**（见下） |
+
+要点：
+
+- **内核 ticker 驱动**，不是前端轮询：没有浏览器、没人登录也会执行。
+- **任务跑在插件自己的 VM 上**，与请求钩子共用 `plugin.state` —— 任务写的
+  token 对下一个请求立即可见，也会经同一套机制持久化到 `.<name>.state.json`。
+- 任务函数出错只记 `hook_errors`，不影响钩子与转发。
+- 没有 `schedule` 的插件不获得任何额外 VM/goroutine（零开销）。
+- **超时不能中断一个运行中的 Lua 调用**（本 LuaJIT 构建的硬限制）。任务超过
+  10 分钟不返回，该插件被**停用**（`disarmed`）：不再调度，钩子快速失败并
+  记 `hook_errors`。这是为了不让一个卡死的任务把请求永久堵在它那把锁后面。
+  修好插件后重新加载（重传/重启）即可恢复。
+
+> 同一条 6.1 的约束：任务占着插件锁，所以**任务里不要写死循环**。
+
+### 5.5 自定义对外挂载点
+
+插件的 HTTP 端点默认挂在 `/api/plugins/http/...` 下。同一个声明**同时**可在两个
+公开命名空间下访问，只要把 `path` 的第一段起成插件名：
+
+```lua
+plugin.http = {
+  { path = "myplugin/callback", method = "POST", handler = "on_callback" },
+}
+```
+
+该端点因此可在三个地方访问：
+
+| 路径 | 用途 |
+|---|---|
+| `/api/plugins/http/myplugin/callback` | 默认前缀（带登录态的管理 UI） |
+| `/plugins/myplugin/callback` | 公开（浏览器路径） |
+| `/v1/plugins/myplugin/callback` | 公开（API 路径） |
+
+可占用的命名空间只有后两个：
+
+| 前缀 | 例子 |
+|---|---|
+| `/plugins/` | `/plugins/myplugin/callback` |
+| `/v1/plugins/` | `/v1/plugins/myplugin/callback` |
+
+限制（都是为了不把内核路由变成可被插件改写的）：
+
+- **不能影子内核路由**：与 `/v1/chat/completions`、`/v1/messages`、`/api/...`
+  等重叠的挂载在加载时就被拒绝，错误信息点名冲突对象。
+- **只能是上面两个命名空间**，不接受任意前缀（否则可以拦截推理请求）。
+- 鉴权沿用端点的 `roles` / `keys` / `public` 语义（见 §5.1）；
+  `/v1/plugins/...` 属于 API 路径，未认证返回 401；`/plugins/...` 为浏览器
+  路径，未认证会跳登录页。
+- 未声明的路径不会因为落在允许命名空间而被放行——不在 handler 表里就是 404。
+
+### 5.6 前缀代理：把插件自起的服务挂到网关同端口
+
+前节的自定义挂载点只能回答**由插件计算出来的内容**（`{status, headers, body/json}`）。
+若插件自己起了一个 TCP/HTTP 服务（典型：服务端浏览器栈 + noVNC），客户端就需要
+**字节级转发**，包括 **WebSocket 升级** —— 前者做不到。
+
+为此声明 `prefix = true`，handler 每次请求返回一个**回环目标**：
+
+```lua
+plugin.http = {
+  { path = "myplugin/proxy/:session", method = "GET", prefix = true,
+    roles = { "admin" }, handler = "proxy" },
+}
+
+function plugin.proxy(req)
+  -- 每次请求都问一遍：会话没了/过期了就返回 404，地址自然失效
+  if not sessionAlive(req.params.session) then
+    return { status = 404, json = { error = "no such session" } }
+  end
+  return { proxy = "http://127.0.0.1:60100" }
+end
+```
+
+规则：
+
+- **只能返回 `{ proxy = <目标> }`**，没有 `status` / `body`。写了 `status` 会被忽略。
+- 目标必须是 **带显式端口的字面回环 IP**（`http://127.0.0.1:60100`）。
+  拒绝 `https://`、`localhost` 这类主机名、无端口、带 userinfo/path/query 的地址。
+  理由：主机名可能被 hosts / DNS / 代理环境改写；通用能力只接受确定性地址。
+- 路由匹配同前缀下的**所有子路径**，`:param` 仍照常捕获，`req.path` 是完整
+  请求路径，子路径由插件自己从 `req.path` 里切。
+- **鉴权先于代理**：`roles` / `keys` / `public` 语义与普通端点一致。
+- **不转发网关凭据**：`Authorization`、`Cookie`、`api_key` 都不会传给后端服务。
+- **拒绝跨站**：`Origin` 与网关自身 host 不一致返回 403。WebSocket 由浏览器
+  发起且自动带 Cookie，否则第三方网页可借管理员会话开隧道。
+- **撤销点唯一**：handler 不再返回目标（或返回 404）即失效，内核不缓存目标。
+
+客户端拿到的 URL 是**相对网关口**的绝对路径（如
+`/api/plugins/http/myplugin/proxy/<id>/vnc.html`），因此远程访问、HTTPS 入口、
+Cookie 鉴权都自然生效，无需暴露额外端口。
+
+### 5.7 在运行时安装/更新适配器（token 轮换）
+
+插件可以在运行时安装、改写**自己的**适配器 —— 这是"带自动轮换的认证"的实现
+方式：拿着新 token 重写适配器源码，后续请求自然使用新逻辑（内核重建该适配器
+的池）。**可直接在插件 `.lua` 里调用**：
+
+```lua
+-- 每小时的定时任务：轮换 token 并重写适配器
+function plugin.refresh_token()
+  local tok = fetch_new_token()          -- 插件自己的取 token 逻辑
+  local name, err = modelrouter.adapters.install(
+    "myauth",
+    'return { name = "myauth", endpoint = "/chat/completions",' ..
+    ' headers = { ["Authorization"] = "Bearer ' .. tok .. '" } }'
+  )
+  if err then log("error", err) end
+end
+```
+
+| 调用 | 参数 | 返回 |
+|---|---|---|
+| `modelrouter.adapters.install(name, source)` | 适配器名 + Lua 源码 | `(生效名, nil)` 或 `(nil, err)` |
+| `modelrouter.adapters.update(name, source)` | 同上（只限自己装的） | `(true, nil)` 或 `(nil, err)` |
+| `modelrouter.adapters.remove(name)` | 适配器名（只限自己装的） | `(true, nil)` 或 `(nil, err)` |
+
+出错时不抛异常，而是返回 `(nil, err)`——与钩子/HTTP 处理器同一套约定，
+一个手滑的插件错误不会带走网关。
+
+安全规则（由 Go 侧强制，不是约定）：
+
+- 插件**只能改自己装的**适配器（归属表在 Go 侧记录，调用身份是插件启动时的
+  固定名字，无法冒充）。
+- 插件**不能影子内置适配器**（否则等于篡改发行物）；运维仍可自行覆盖。
+- 名字经过路径安全校验（无穿越/绝对路径/空白）。
+
+> 这条能力配合 §5.4 的定时任务，就是完整的自动轮换：**没有浏览器、没有人
+> 操作时，插件按时间自己把 token 换掉**。
+
 ---
 
 ## 6. 运行时约束（重要）
 
-### 6.1 一个插件 = 一个 Lua 状态
+### 6.1 一个插件 = 一个 Lua 状态（也就是一个 VM）
 
 适配器可以有多个独立的 worker 状态（它们无状态，这是对的）。**插件不行**：
 钩子通常往 `plugin.state` 里累加，多个状态就意味着总额被劈成几份，而写价格只
 写进其中一个状态，钩子恰好跑到另一个时**所有请求按 0 计费**。
 
-因此插件持有**唯一一个**状态，所有入口（`Fire` / `State` / `SetState`）用互斥锁
+因此插件持有**唯一一个**状态，所有入口（`Fire` / `State` / `SetState` / **定时任务**）
-串行化。代价：**一个卡住的钩子会卡住所有插件的钩子**。所以——
+都用这同一把互斥锁串行化。
 
-> **钩子必须短、同步、不阻塞。** 不要在里面做 HTTP 请求、睡眠或重计算。
+注意： 这条约束对**定时任务尤其关键**：任务和钩子必须是同一个 VM，否则任务写的
+`plugin.state` 是一个请求路径永远读不到的副本 —— token 轮换看起来成功了，
+对实际流量完全无效。
+
+代价：**一个卡住的钩子会卡住该插件自己的钩子**（跨插件仍然并行，每插件各自
+一把锁、各自一个 VM）。所以——
+
+> **钩子与定时任务都必须短、同步、不阻塞。** 不要在钩子里做 HTTP 请求、睡眠
+> 或重计算；长耗时的周期工作应放进 `PLUGIN.schedule`（见 §5.4），它在自己的
+> 调度 goroutine 上跑，但仍然持有这把锁，因此不能是死循环。
 
 ### 6.2 异常被隔离
 
@@ -326,28 +519,70 @@ PUT    /api/plugins/{name}/state     替换状态（admin）
 与适配器相同：`json.encode` / `json.decode`、`log(level, msg)`、
 `hmac_sha256_hex` / `sha256_hex` / `base64_encode` / `tohex`。
 
-⚠️ `os.date` 在极简运行时里可能不可用（计费插件因此有降级路径）。
+注意： `os.date` 在极简运行时里可能不可用（计费插件因此有降级路径）。
 
 ### 6.4 命名
 
-- 全局 `__llmsproxy_plugin` 存放插件返回的 table，**不要占用**
+- 全局 `__llmsproxy_plugin` 是内核存放插件返回 table 的**内部位置**，不属于 SDK，**不要占用**
 - 插件间状态完全隔离，一个插件改不了另一个的全局
 
 ---
 
-## 7. 内置计费插件
+## 7. Billing 插件示例
 
-`plugins/billing.lua`，默认随首启 seed 进来，开箱可用。
+`internal/lua/plugins/billing.lua` 是仓库中的可选插件示例，不随核心安装包分发，也不会在启动时自动安装。需要使用计费功能时，应将该脚本单独复制到配置的 `plugin_dir` 后再重启网关。
 
-### 7.1 计费维度
+### 7.0 分层：billing 完全在插件里，内核没有任何 billing 代码
 
-| 维度 | 用途 | 优先级 |
+`billing` 是**插件自己的事**：规则文件、编译、计价、账目、规则编辑页都由
+`plugins/billing.lua` 负责。内核里没有 billing 的配置结构、端点、编译或迁移逻辑
+—— 它只提供所有插件通用的能力（加载 `.lua`、插件状态持久化、插件目录、
+`plugin.http` 路由、插件可用的 `io`/`os`）。
+
+| 项 | 位置 |
+|---|---|
+| 规则文件 | `<plugin_dir>/billing.json`（插件自己读写；内核不知道这个文件） |
+| 计费账目 | 插件状态文件 `<plugin_dir>/.billing.state.json`（内核的通用状态持久化） |
+| 规则编辑 API | `/api/plugins/http/billing/rules`（GET/PUT/POST/DELETE，插件路由） |
+| 编译/计价 | 插件内 `compileRules()` / `priceFor()` / `costFor()` |
+
+规则以**源名**（`source`）为键，而不是 URL：URL→源名的映射属于内核配置，插件
+不需要为了它和内核对接；编辑页在浏览器里，直接取 `/api/sources` 给运营选源名。
+`url: "*"` 只表示"全局模型价目表"（模型维度本来就不分源）。
+
+> 历史：早期 `billing:` 写在 `config.yaml`、编译在 Go 的 `internal/billing` 里。
+> GitCode issue #1 的根因正是"Go 与 Lua 各写一半覆盖语义"，组合起来把模型价
+> 静默清零。现在内核侧 billing 代码已全部删除，规则文件也由插件自持。
+
+### 7.1 计费维度### 7.1 计费维度
+
+| 维度 | 用途 | token 价覆盖顺序 |
 |---|---|---|
-| `sources` | 源的固定价（如按请求计费的转发服务） | 中 |
+| `default` | 兜底价（通常全 0） | 最低 |
-| `models` | 模型的 per-token 价 | 高 |
+| `models` | 模型的 per-token 价 | 中 |
+| `sources` | 源级价：固定价，或覆盖该源全部模型的 token 价 | 高 |
 | `keys` | 单个网关 key 的覆盖价 | **最高** |
 
-**token 价**优先级：`keys` > `models` > `default`。
+**token 价**优先级：`keys` > `sources` > `models` > `default`。
+（实现顺序就是"后写覆盖前写"：`default` → `models` → `sources` → `keys`；
+所以**源级 token 价会覆盖模型价**。这是有意的：运营要能给某个源整体改价。）
+
+⚠️ 正因为 `sources` 会覆盖 `models`：**`mode: token` 的规则不产生任何源条目**。
+规则未列出的模型会进入未定价统计（可见），而不是静默计 0 —— 这是 issue #1 / PR #2
+的收口语义。只有 `free` / `subscription`（"边际成本为零"）会写有意为之的
+`{prompt:0, completion:0}` 零价条目。
+
+规则写法：
+
+```json
+{"active":"current","profiles":[{"id":"current","default":"unpriced","rules":[
+  {"source":"localzen","mode":"free"},
+  {"source":"codearts","mode":"token",
+   "models":{"glm-5.3":{"prompt":"0.15","completion":"0.60"}}},
+  {"url":"*","mode":"token","models":{"some-model":{"prompt":"1.0","completion":"2.0"}}}
+]}]}
+```
+
 **`per_request` 固定价是叠加的**（不覆盖），所以一个生图模型可以既算 token 又收固定费。
 **`cache_discount`** 见 §7.7。
 
@@ -413,7 +648,7 @@ curl -H "Authorization: Bearer $ADMIN_KEY" \
 { "prices": { ... }, "count_failures": true }
 ```
 
-### 7.5 ⚠️ 计费插件只报表，不执法
+### 7.5 注意： 计费插件只报表，不执法
 
 **网关自己的配额会计（`internal/gateway/stats.go`，入口处强制）才是限额权威。**
 本插件不参与任何路由或配额决策。
@@ -435,6 +670,17 @@ curl -H "Authorization: Bearer $ADMIN_KEY" \
 
 仪表盘上有 "Unpriced" 卡片。**这个数应该是 0**；不是 0 就去补价目。
 
+**profile 级兜底**：`profiles[].default` 可设为 `free` 或 `unpriced`（默认）。
+它是插件运行时的策略（编译期写入 `prices.default_mode`）：
+
+| `default` | 没有单独规则的源 | `unpriced_reqs` |
+|---|---|---|
+| `unpriced`（省略时） | 零价，计入未定价债务 | +1 |
+| `free` | 零价，视为**刻意免费** | 不增加 |
+
+显式 `mode: unpriced` 的源即使在 `default: free` 下也仍然计入（"只统计不收费"
+的声明不能被兜底策略吞掉）。
+
 注意"未定价"不等于"免费"：这些请求的 `requests` / token 数**照常计入**
 `total` 与各维度，只有金额是 0。
 
@@ -487,7 +733,7 @@ cached = cache_hit_tokens                   → 全价 × cache_discount
 插件的 `log(level, msg)` 输出到网关的日志：
 
 ```
-journalctl -u llmsproxy -f | grep -i plugin
+journalctl -u modelrouter -f | grep -i plugin
 ```
 
 ---
@@ -497,10 +743,12 @@ journalctl -u llmsproxy -f | grep -i plugin
 | | 适配器 | 插件 |
 |---|---|---|
 | 位置 | `adapter_dir` | `plugin_dir` |
-| 作用 | 上游协议转换 | 请求流水线 + 界面 |
+| 作用 | 上游协议转换 | 请求流水线 + 界面 + 周期任务 |
-| 钩子 | `transform_request` / `transform_response` / `transform_stream_chunk` / `build_headers` / `transform_error` | `request_start` / `routed` / `request_end` |
+| 钩子 | `transform_request` / `transform_response` / `transform_stream_chunk` / `build_headers` / `transform_error` / `serialize_*` | `request_start` / `routed` / `request_end` |
-| 状态 | 每个源独立，多 worker | 单状态，见 §6.1 |
+| 周期任务 | **无**（无状态请求级） | `PLUGIN.schedule`（内核 ticker 驱动） |
-| 出错后果 | 该 (源,模型) 进入冷却退避 | 仅该功能缺失 |
+| 对外端点 | 不开放（见 §5.5） | `/api/plugins/http/...` + 自定义挂载 |
+| 状态 | 每个源独立，多 worker | **单 VM、单状态**，见 §6.1 |
+| 出错后果 | 该 (源,模型) 进入冷却退避 | 仅该功能缺失；卡死则停用该插件 |
 | 鉴权 | 需要 `gateway_keys` 之外的独立凭据 | 网关 key |
 | 必需性 | 核心 | 可选，缺了网关照常跑 |
 

docs/routing.md

@@ -0,0 +1,111 @@
+# 路由与调度
+
+> 本文档由 README 拆分而来：模型路由、AUTO 链、冷却与半冷却、源与适配器、源模板。
+> 返回 [README](../README.md)。
+
+## 模型路由
+
+`/v1/chat/completions` 的 `model` 解析顺序：
+
+1. `source/model` 或 `source:model` 前缀 → 指定源；
+2. 精确匹配某个源的 `model`；
+3. 配置 `default_model: AUTO` 时 → 走优先级页保存的 AUTO 链（见下）；
+4. 否则回落（返回错误）。
+
+任何 OpenAI 客户端，只要 `model` 设为某个源的 `name/任意名`，即可锁定走该源；
+设为 `AUTO`（或网关配了 `default_model: AUTO`）即自动按优先级链选源。
+
+## AUTO 链（优先级页）
+
+AUTO 调度**只**由优先级页保存的规则决定，源配置里的 `models[].priority` 数字不再
+参与调度、也不再显示。**聊天与生图是两套独立的链**，在优先级页面通过「聊天 / 生图」
+按钮切换编辑：
+
+- **聊天链**（`runtime_file` 的 `auto` 字段）：服务 `POST /v1/chat/completions`
+  的 `model: AUTO` 请求；只接受 `kind: chat` 模型，配置在链里的生图槽位会被忽略。
+- **生图链**（`runtime_file` 的 `auto_image` 字段）：服务 `POST /v1/images/generations`
+  的 `model: AUTO` 请求；只接受 `kind: image` 模型。未配置生图链时，AUTO 生图回落
+  到「所有含生图模型的源、按注册顺序」的旧行为，兼容旧配置。
+
+每条规则 = 一个「槽位」：`{ model, source, tier, token_quota, period, hours }`。
+
+- `tier` 表示优先级档位：同一档的模型并排、共享该优先级；档位从上到下递减。
+- 同一模型可配置多个槽位（如 A 源低配 → B 源低配 → A 源高配 → B 源高配），按档位顺延。
+- `token_quota` > 0 时该槽位在重置周期内用满即顺延到下一槽位；`period` 支持
+  `hour` / `week` / `month` / `nhour`（配合 `hours`），空 = 不限。
+- 同 tier 内按「偏好分」排序，失败槽位冷却后自动跳过；冷却 / 配额耗尽 / hard
+  错误会顺延到下一 tier，全部失败时返回 503 并附上每档失败摘要。
+
+## 冷却与半冷却探测（自愈调度）
+
+槽位失败后进入指数退避冷却（5s → 10s → 20s …，上限 **5 分钟**）。冷却不是全时段
+硬闸：
+
+- **前半段完全静默**：刚失败的槽位不接任何流量，避免对故障上游持续施压。
+- **后半段放行 1 个探测**：窗口过半后，同一 `(源, 模型)` 最多允许 **一个** 请求作为
+  探测通过。探测槽数 = `clamp(max_concurrent / 10, 1, 2)`——`max_concurrent: 10`
+  的源恰好放行 1 个探测请求。
+- **探测排在最后**：同 tier 内探测候选永远排在健康候选之后，只有没有正常槽可用时
+  才承接真实流量，因此探测不会抢走可用容量。
+- **探测即真实请求**：成功则立即清零失败计数与冷却，槽位当场回到正常轮询——
+  上游恢复/额度恢复后不必等满整个冷却窗口。失败则重开一个新窗口，下次探测顺延到
+  新窗口的中点，不会连续重试。
+
+分类冷却策略：
+
+| 情况 | 冷却 | 说明 |
+|---|---|---|
+| 传输错误 / 5xx | 5s→10s→20s…，上限 5min | 指数退避 |
+| 401 / 403 凭据错误 | 10min | 换 key 后由下一次探测自动接回 |
+| 429 限流 | 30s 固定 | 「太快」不等于「坏了」，不进指数阶梯 |
+| 配额耗尽（`insufficient_quota` / `余额不足` 等） | 对齐配额窗口，上限 30min | **不进指数阶梯、不涨失败计数**，额度恢复即可用 |
+
+WebUI「优先级」页的槽位健康标签会区分 `冷却` / `待探测` / `探测中`，悬停可看窗口
+起点、探测放行时刻与完全恢复时刻。
+
+## 源（Source）与适配器（Adapter）的关系
+
+- **源** 是到某个上游的连接描述：`name`、`base_url`、`api_key`、模型列表与优先级。
+- **适配器** 是协议转换逻辑（Lua 脚本）：把统一 OpenAI 格式请求转成上游原生格式，
+  再把上游响应/流式分块转回统一格式。
+- 一个适配器可被多个源复用（如 `openai.lua` 同时服务多个 OpenAI 兼容站点）；
+  同一个源也可以切换不同适配器（改 `adapter` 字段即可）。
+- 源决定“连谁、暴露哪些模型”，适配器决定“怎么对话”——二者在 `sources[]` 条目中
+  通过 `adapter` 字段关联。
+
+加载流程（`internal/core` 的 `Core` 负责装配）：
+
+1. 启动时 `lua.NewVM(adapter_dir)` 加载全部适配器：内置适配器（编译期 embed）+
+   `adapter_dir` 下的同名覆盖文件；内置文件写在代码内，覆盖文件要求更高优先级。
+2. `config.Load` 读取 `config.yaml`，`config.NewStore(runtime_file)` 读 WebUI 改动的
+   运行时源，二者按名称合并成完整源列表。
+3. `rebuildRegistry` 为每个源创建 `provider.Provider`（持有目标适配器），并按源的
+   并发上限配置适配器 worker 池**上限**（不预分配，见下）。
+4. 请求进来时 `Registry` 按 `model` 路由到 Provider，Provider 调适配器
+   `transform_request` → HTTP 发送 → `transform_response` / `transform_stream_chunk`。
+
+WebUI 上的"新增/编辑源"、"上传 Lua 适配器"、"改 AUTO 优先级链"与"密钥模型范围"
+都即时生效（写入运行时文件或 `adapter_dir` 后重新装配，无需重启）；直接编辑
+`config.yaml` / `adapter_dir` 下的文件则需要重启进程才会重新加载。
+
+在 WebUI 删除源/适配器时：运行时（WebUI 创建）的源与 `.lua` 文件会真正移除；
+`config.yaml` 中定义的基源无法改写配置文件，采用删除标记隐藏（重启后仍隐藏），
+在 UI 里重新添加同名源即可恢复。
+
+## 源模板（多 key 负载均衡）
+
+多个上游 key 共享同一套配置（URL / 适配器 / 模型列表 / 并发 / rpm）时，逐一复制
+整段 source 太冗余。**源模板**保存除 `name` 和 `api_key` 以外的全部源字段，
+从模板一键创建多个仅 key 不同的源，天然支持多 key 负载均衡（每个展开源独立调度、
+独立冷却、独立健康状态）。
+
+- WebUI「源」页右上角「**模板管理**」按钮：列出全部模板，可编辑 / 删除 / 新建。
+- 添加源弹窗右上角两个按钮：
+  - **从模板创建** → 选模板 → 表单自动填充（name + key 仍需你填）；
+  - **存为模板** → 把当前表单非 key/name 字段存为新模板。
+- 模板持久化在 `runtime_file` 的 `source_templates` 字段，与运行时源一样热管理，
+  改完即时生效。模板本身不含密钥，可以安全共享 / 纳管。
+
+模板不会自己变成源——它只是「配方」。真正承担流量的是从模板创建出来的、带 key
+的具体源。
+

docs/webui.md

@@ -0,0 +1,23 @@
+# WebUI
+
+> 本文档由 README 拆分而来。
+> 返回 [README](../README.md)。
+
+## WebUI
+
+内置管理界面（`GET /`），登录后可在浏览器完成：
+
+- **状态页**：源在线状态（实时探测 + 悬停看错误）、模型/源/key 用量统计、请求记录，
+  支持按时间范围导出 CSV；点击模型可生成 pin 到该模型的连接配置。
+- **对话页**：流式/非流式调试。
+- **密钥页**：创建/编辑网关 key，为每个 key 配模型范围（模型 + 源 + token 配额 + 周期），
+  管理员管理全部 key，用户只看到自己的 key。每个模型砖块显示自己的配额徽标
+  （如 `1.0K 77×·1h`，未设配额显示 `∞`），点开可编辑该模型的 token 配额、
+  请求数配额与重置周期 —— 配额按模型独立生效，一个用满不影响同一 key 的其它模型。
+  「我的密钥」页对用户展示本 key 的模型范围与各自配额。
+- **优先级页**：拖拽积木配置 AUTO 链档位。
+- **源页**：在线增删改上游源（API key 等敏感字段加密落盘）。
+- **适配器页**：上传 / 删除 Lua 适配器脚本。
+
+改动写入 `runtime_file`（重启仍生效）。
+

e2e/adapter_takeover_test.go

@@ -0,0 +1,286 @@
+package e2e
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// 本文件端到端验证 R2/R5 的核心主张：**对外服务的 wire 格式由适配器决定，
+// 内核只调度**；适配器不声明时内核的 OpenAI 兜底仍然可用。
+//
+// 与其他测试的分工：
+//   - internal/gateway/messages_test.go 在**进程内**验证同一件事（起 core+gateway，
+//     不打真二进制）。它跑得快，但绕过了一些只在真实进程里发生的事：Lua 适配器的
+//     磁盘 seed、配置加载、systemd 式启动路径、真 HTTP 栈。
+//   - 本文件打**真编译出的二进制**，用一个只说 Anthropic Messages 的假上游，
+//     证明整条链路（真进程 → 真 Lua 适配器 → 真上游 → 适配器序列化 → 客户端）
+//     确实成立，而不是只在进程内装配下成立。
+//
+// 这是用户明确问到的"适配器接手对外服务的部分"的验收点。
+
+// anthropicMockUpstream 是只懂 Anthropic Messages 协议的假上游。
+//
+// 它会对**不符形状的请求主动报错**，而不是宽容接受 —— 宽容的 mock 会让
+// "内核把 OpenAI 形状发给了 Anthropic 上游"这种错误测试通过。
+func anthropicMockUpstream(t *testing.T) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Anthropic 的请求路径必须是 <base>/messages
+		if !strings.HasSuffix(r.URL.Path, "/messages") {
+			t.Errorf("请求打到了非 Anthropic 路径: %s", r.URL.Path)
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		body, _ := io.ReadAll(r.Body)
+		var probe map[string]interface{}
+		if err := json.Unmarshal(body, &probe); err != nil {
+			t.Errorf("上游收到非 JSON 请求体: %s", body)
+			w.WriteHeader(http.StatusBadRequest)
+			return
+		}
+		// Anthropic 形状的硬性要求：messages 是数组；不能有 OpenAI 的 choices 概念
+		if _, ok := probe["messages"]; !ok {
+			t.Errorf("上游请求体缺 messages: %s", body)
+		}
+		if _, ok := probe["max_tokens"]; !ok {
+			t.Errorf("Anthropic 要求 max_tokens，请求体缺此项: %s", body)
+		}
+		// 反向断言：OpenAI 兜底字段不该出现在 Anthropic 上游请求里
+		for _, k := range []string{"choices", "chat_completions"} {
+			if _, bad := probe[k]; bad {
+				t.Errorf("Anthropic 上游请求体混入了 OpenAI 字段 %q: %s", k, body)
+			}
+		}
+
+		if stream, _ := probe["stream"].(bool); stream {
+			w.Header().Set("Content-Type", "text/event-stream")
+			fl, _ := w.(http.Flusher)
+			frames := []string{
+				`event: message_start` + "\n" + `data: {"type":"message_start","message":{"id":"msg_e2e","type":"message","role":"assistant","model":"claude-e2e","content":[],"stop_reason":null,"usage":{"input_tokens":3,"output_tokens":0}}}`,
+				`event: content_block_start` + "\n" + `data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+				`event: content_block_delta` + "\n" + `data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"pong"}}`,
+				`event: content_block_stop` + "\n" + `data: {"type":"content_block_stop","index":0}`,
+				`event: message_delta` + "\n" + `data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":1}}`,
+				`event: message_stop` + "\n" + `data: {"type":"message_stop"}`,
+			}
+			for _, f := range frames {
+				fmt.Fprint(w, f+"\n\n")
+				if fl != nil {
+					fl.Flush()
+				}
+			}
+			return
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"id":"msg_e2e","type":"message","role":"assistant","model":"claude-e2e",`+
+			`"content":[{"type":"text","text":"pong"}],"stop_reason":"end_turn",`+
+			`"usage":{"input_tokens":3,"output_tokens":1}}`)
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+// writeAnthropicConfig 写一份把源挂到 anthropic 适配器上的配置。
+func writeAnthropicConfig(t *testing.T, dir, listenAddr, upstreamURL string) string {
+	t.Helper()
+	cfg := "listen: " + listenAddr + "\n" +
+		"gateway_keys:\n  - sk-e2e-0001\n" +
+		"default_model: claude-e2e\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"sources:\n" +
+		"  - name: anth\n" +
+		"    base_url: " + upstreamURL + "\n" +
+		"    adapter: anthropic\n" +
+		"    api_key: test-key\n" +
+		"    max_concurrent: 32\n" +
+		"    models:\n" +
+		"      - id: claude-e2e\n" +
+		"        kind: chat\n" +
+		"        priority: 100\n"
+	path := filepath.Join(dir, "config.yaml")
+	if err := os.WriteFile(path, []byte(cfg), 0o644); err != nil {
+		t.Fatalf("write config: %v", err)
+	}
+	return path
+}
+
+// TestAdapterTakeoverNonStream：非流式，Anthropic 客户端形状进、出。
+func TestAdapterTakeoverNonStream(t *testing.T) {
+	up := anthropicMockUpstream(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	cfg := writeAnthropicConfig(t, dir, addr, up.URL)
+	bin := buildBinary(t)
+	g := startGateway(t, bin, addr, cfg)
+
+	resp, body := g.do("POST", "/v1/messages",
+		`{"model":"claude-e2e","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`, true)
+	if resp == nil || resp.StatusCode != 200 {
+		t.Fatalf("status=%v body=%q", statusOf(resp), body)
+	}
+	// 出向必须是 Anthropic Messages 形状，不能是 OpenAI choices 形状
+	if strings.Contains(body, `"choices"`) {
+		t.Errorf("Anthropic 响应里混入了 OpenAI choices 外壳: %s", body)
+	}
+	var out map[string]interface{}
+	if err := json.Unmarshal([]byte(body), &out); err != nil {
+		t.Fatalf("响应不是 JSON: %v: %s", err, body)
+	}
+	if out["type"] != "message" {
+		t.Errorf(`响应 type = %v, 期望 "message"（Anthropic 形状）: %s`, out["type"], body)
+	}
+	// 文本内容必须能取出来 —— 这是"内容确实穿过适配器回来了"的证据
+	if !strings.Contains(body, "pong") {
+		t.Errorf("响应里没有上游返回的文本: %s", body)
+	}
+	if _, ok := out["usage"]; !ok {
+		t.Errorf("Anthropic 响应应带 usage: %s", body)
+	}
+}
+
+// TestAdapterTakeoverStream：流式也要是 Anthropic SSE 序列，
+// 且**不得**出现 OpenAI 的 [DONE] 终止标记（R2 的"不混合"）。
+func TestAdapterTakeoverStream(t *testing.T) {
+	up := anthropicMockUpstream(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	cfg := writeAnthropicConfig(t, dir, addr, up.URL)
+	bin := buildBinary(t)
+	g := startGateway(t, bin, addr, cfg)
+
+	resp, body := g.do("POST", "/v1/messages",
+		`{"model":"claude-e2e","max_tokens":64,"stream":true,"messages":[{"role":"user","content":"hi"}]}`, true)
+	if resp == nil || resp.StatusCode != 200 {
+		t.Fatalf("stream status=%v body=%q", statusOf(resp), body)
+	}
+	if !strings.Contains(body, "message_start") {
+		t.Errorf("Anthropic 流应以 message_start 开头: %s", body)
+	}
+	if !strings.Contains(body, "message_stop") {
+		t.Errorf("Anthropic 流应以 message_stop 收尾: %s", body)
+	}
+	if strings.Contains(body, "[DONE]") {
+		t.Errorf("Anthropic 流里出现了 OpenAI 的 [DONE]: %s", body)
+	}
+	if strings.Contains(body, "chat.completion.chunk") {
+		t.Errorf("流里混入 OpenAI chunk 形状: %s", body)
+	}
+	if !strings.Contains(body, "pong") {
+		t.Errorf("流里没有上游文本: %s", body)
+	}
+}
+
+// TestAdapterTakeoverStress 是对外服务路径的真实压力测试：
+// 并发打真二进制，混合流式/非流式，验证无 5xx、上游命中数正确、
+// 且每个响应都是完整合法的 Anthropic 形状（适配器在并发下不串味）。
+func TestAdapterTakeoverStress(t *testing.T) {
+	up := anthropicMockUpstream(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	cfg := writeAnthropicConfig(t, dir, addr, up.URL)
+	bin := buildBinary(t)
+	g := startGateway(t, bin, addr, cfg)
+
+	// max_concurrent 配成 32，让 96 个请求基本都能真正流过适配器 —— 这个测试
+	// 要压的是**并发下序列化是否串味**，不是内核的容量拒绝策略。真的撞上上限
+	// 时收到的 429 也是合法背压（见下方断言），不算失败。
+	const (
+		workers   = 12
+		perWorker = 8
+	)
+
+	var (
+		wg      sync.WaitGroup
+		mu      sync.Mutex
+		ok      int
+		shed    int
+		failed  []string
+		streams int
+	)
+	start := make(chan struct{})
+
+	for w := 0; w < workers; w++ {
+		wg.Add(1)
+		go func(w int) {
+			defer wg.Done()
+			<-start
+			for i := 0; i < perWorker; i++ {
+				isStream := (w+i)%2 == 0
+				var payload string
+				if isStream {
+					payload = `{"model":"claude-e2e","max_tokens":32,"stream":true,"messages":[{"role":"user","content":"hi"}]}`
+				} else {
+					payload = `{"model":"claude-e2e","max_tokens":32,"messages":[{"role":"user","content":"hi"}]}`
+				}
+				resp, body := g.do("POST", "/v1/messages", payload, true)
+				mu.Lock()
+				switch {
+				case resp == nil:
+					failed = append(failed, "nil response")
+				// 429 = 内核的容量背压（provider busy）。它是**正确**行为，
+				// 不是缺陷：并发超限时就该如实拒绝，而不是排队或伪造响应。
+				// 只要求它是结构良好的错误，不要求 200。
+				case resp.StatusCode == http.StatusTooManyRequests:
+					if !strings.Contains(body, "busy") && !strings.Contains(body, "error") {
+						failed = append(failed, "429 不是结构化错误: "+body)
+					} else {
+						shed++
+					}
+				// 5xx 才是真失败 —— 那是内核或适配器崩了。
+				case resp.StatusCode >= 500:
+					failed = append(failed, fmt.Sprintf("5xx status=%d body=%.120s", resp.StatusCode, body))
+				case resp.StatusCode != 200:
+					failed = append(failed, fmt.Sprintf("status=%d body=%.120s", resp.StatusCode, body))
+				case !strings.Contains(body, "pong"):
+					// 200 但内容不对：说明并发下适配器把响应串味了
+					failed = append(failed, "missing upstream text: "+body)
+				case isStream && !strings.Contains(body, "message_stop"):
+					failed = append(failed, "stream missing message_stop")
+				case !isStream && strings.Contains(body, `"choices"`):
+					failed = append(failed, "openai shell leaked: "+body)
+				default:
+					ok++
+					if isStream {
+						streams++
+					}
+				}
+				mu.Unlock()
+			}
+		}(w)
+	}
+	close(start)
+	wg.Wait()
+
+	if len(failed) > 0 {
+		t.Fatalf("%d/%d 请求失败，前 5 例:\n%s", len(failed), workers*perWorker,
+			strings.Join(failed[:min(5, len(failed))], "\n"))
+	}
+	// 至少大部分请求必须真正流过 200 —— 全被背压拒掉说明并发上限配得太低，
+	// 那这个压力测试就没压到适配器并发。
+	if ok < workers*perWorker/2 {
+		t.Fatalf("只有 %d/%d 请求真正成功，压力未触及适配器并发路径",
+			ok, workers*perWorker)
+	}
+	t.Logf("成功 %d（其中流式 %d），背压 429 %d —— 无 5xx、无 OpenAI 形状泄漏",
+		ok, streams, shed)
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+var _ = time.Second

e2e/conflict_arbitration_e2e_test.go

@@ -0,0 +1,261 @@
+package e2e
+
+import (
+	"encoding/json"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// writeE2EPlugin 落一个插件文件到 e2e 的 plugin_dir（SeedBundled 见空目录才
+// 落地内置插件，所以先放自定义插件再启动网关）。
+func writeE2EPlugin(t *testing.T, dir, name, code string) {
+	t.Helper()
+	p := filepath.Join(dir, "plugins", name)
+	if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(p, []byte(code), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// e2eConfig 返回一份最小 e2e 配置文本（含 plugin_dir）。
+func e2eConfig(dir, listen string) string {
+	return "listen: " + listen + "\n" +
+		"gateway_keys:\n  - sk-e2e-0001\n" +
+		"default_model: AUTO\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"plugin_dir: " + filepath.Join(dir, "plugins") + "\n" +
+		"sources: []\n"
+}
+
+// TestConflictArbitrationE2E（用户要求：服务路径冲突不崩溃不拒启 + WebAPI 裁决）
+//
+// 真二进制全链路：
+//  1. 两个插件声明同一条 /api/plugins/http/ 路径 → 两插件都 loaded，启动正常
+//  2. 请求该路径 → 409，错误信息点名两个候选插件（不静默选边）
+//  3. GET /api/plugins/conflicts → 列出冲突（路径 + 候选 + 未裁决）
+//  4. POST /api/plugins/conflicts 裁决 → 持久化
+//  5. 再请求该路径 → 200，赢家应答（{"from":"B"}）
+//  6. /v1/models 等既有路由全程不受影响（旧通路无回归）
+func TestConflictArbitrationE2E(t *testing.T) {
+	bin := buildBinary(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+
+	writeE2EPlugin(t, dir, "plugA.lua", `
+local P = {}
+P.name = "plugA"
+P.version = "1"
+P.exports = {}
+P.http = { { path = "contested", method = "POST", handler = "h", roles = { "admin" } } }
+function P.h(req) return { status = 200, json = { from = "A" } } end
+return P
+`)
+	writeE2EPlugin(t, dir, "plugB.lua", `
+local P = {}
+P.name = "plugB"
+P.version = "1"
+P.exports = {}
+P.http = { { path = "contested", method = "POST", handler = "h", roles = { "admin" } } }
+function P.h(req) return { status = 200, json = { from = "B" } } end
+return P
+`)
+
+	cfgPath := filepath.Join(dir, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte(e2eConfig(dir, addr)), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	g := startGateway(t, bin, addr, cfgPath)
+
+	// ① 两个插件都加载成功（冲突不拒载）
+	resp, body := g.do(http.MethodGet, "/api/plugins", "", true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("GET /api/plugins: %d %s", resp.StatusCode, body)
+	}
+	var list struct {
+		Plugins []struct {
+			Name   string `json:"name"`
+			Loaded bool   `json:"loaded"`
+		} `json:"plugins"`
+	}
+	if err := json.Unmarshal([]byte(body), &list); err != nil {
+		t.Fatal(err)
+	}
+	loaded := map[string]bool{}
+	for _, p := range list.Plugins {
+		loaded[p.Name] = p.Loaded
+	}
+	if !loaded["plugA"] || !loaded["plugB"] {
+		t.Fatalf("冲突的两个插件都应 loaded: %v", loaded)
+	}
+
+	// ② 请求冲突路径 → 409 + 点名候选（不静默选边）
+	resp, body = g.do(http.MethodPost, "/api/plugins/http/contested", "", true)
+	if resp.StatusCode != http.StatusConflict {
+		t.Fatalf("未裁决冲突路径应 409, 实得 %d (%s)", resp.StatusCode, body)
+	}
+	if !strings.Contains(body, "plugA") || !strings.Contains(body, "plugB") {
+		t.Fatalf("409 应点名两个候选: %s", body)
+	}
+
+	// ③ 冲突列表可见
+	resp, body = g.do(http.MethodGet, "/api/plugins/conflicts", "", true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("GET conflicts: %d %s", resp.StatusCode, body)
+	}
+	var cl struct {
+		Routes []struct {
+			Path       string   `json:"path"`
+			Candidates []string `json:"candidates"`
+			Resolved   string   `json:"resolved"`
+		} `json:"routes"`
+	}
+	if err := json.Unmarshal([]byte(body), &cl); err != nil {
+		t.Fatal(err)
+	}
+	if len(cl.Routes) != 1 || cl.Routes[0].Path != "contested" || cl.Routes[0].Resolved != "" {
+		t.Fatalf("冲突形态不对: %+v", cl)
+	}
+
+	// ④ 裁决 plugB 赢
+	arbit, _ := json.Marshal(map[string]string{
+		"kind": "route", "path": "contested", "winner": "plugB",
+	})
+	resp, body = g.do(http.MethodPost, "/api/plugins/conflicts", string(arbit), true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("POST conflicts 裁决: %d %s", resp.StatusCode, body)
+	}
+
+	// ⑤ 再请求 → 200 且赢家应答
+	resp, body = g.do(http.MethodPost, "/api/plugins/http/contested", "", true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("裁决后请求冲突路径应 200, 实得 %d (%s)", resp.StatusCode, body)
+	}
+	if !strings.Contains(body, `"from":"B"`) {
+		t.Fatalf("赢家 plugB 应答, 实得 %s", body)
+	}
+
+	// ⑥ 旧通路无回归：/v1/models 正常
+	resp, _ = g.do(http.MethodGet, "/v1/models", "", true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("旧通路 /v1/models 应 200, 实得 %d", resp.StatusCode)
+	}
+}
+
+// TestAdapterConflictE2E：插件装与内置同名适配器 → 挂起（现任保留），
+// /api/plugins/conflicts 列出，裁决 candidate 后插件版生效。
+func TestAdapterConflictE2E(t *testing.T) {
+	bin := buildBinary(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+
+	writeE2EPlugin(t, dir, "shadow.lua", `
+local P = {}
+P.name = "shadow"
+P.version = "1"
+P.exports = { "install" }
+function P.install(args)
+  local src = 'local a={} a.name="openai" a.version="SHADOW_V1" a.endpoint="/x" return a'
+  local name, err = __llmsproxy_adapters.install("openai", src)
+  return { name = name, err = err }
+end
+return P
+`)
+	cfgPath := filepath.Join(dir, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte(e2eConfig(dir, addr)), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	g := startGateway(t, bin, addr, cfgPath)
+
+	// ① 安装 → 返回 err（pending 语义），内置版本不变
+	resp, body := g.do(http.MethodPost, "/api/plugins/shadow/call/install", "{}", true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("call install: %d %s", resp.StatusCode, body)
+	}
+	// Call 端点把插件返回值包在 {"result": …} 里；挂起语义的断言在 result 层做。
+	var wrapped struct {
+		Result struct {
+			Name *string `json:"name"`
+			Err  string  `json:"err"`
+		} `json:"result"`
+	}
+	if err := json.Unmarshal([]byte(body), &wrapped); err != nil {
+		t.Fatal(err)
+	}
+	out := wrapped.Result
+	if out.Name != nil {
+		t.Fatalf("挂起时不得返回生效名: %s", body)
+	}
+	if !strings.Contains(out.Err, "openai") || !strings.Contains(out.Err, "conflict") {
+		t.Fatalf("错误应点名冲突: %q", out.Err)
+	}
+
+	// ② 冲突列表可见
+	resp, body = g.do(http.MethodGet, "/api/plugins/conflicts", "", true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("GET conflicts: %d %s", resp.StatusCode, body)
+	}
+	var cl struct {
+		Adapters []struct {
+			Name      string `json:"name"`
+			Candidate string `json:"candidate"`
+		} `json:"adapters"`
+	}
+	if err := json.Unmarshal([]byte(body), &cl); err != nil {
+		t.Fatal(err)
+	}
+	if len(cl.Adapters) != 1 || cl.Adapters[0].Name != "openai" || cl.Adapters[0].Candidate != "shadow" {
+		t.Fatalf("冲突形态不对: %+v", cl)
+	}
+
+	// ③ 适配器列表里 openai 仍是内置版本（现任继续服务，不被静默覆盖）
+	assertOpenAIVersion := func(want string, msg string) {
+		t.Helper()
+		resp, body := g.do(http.MethodGet, "/api/adapters", "", true)
+		if resp.StatusCode != 200 {
+			t.Fatalf("GET /api/adapters: %d %s", resp.StatusCode, body)
+		}
+		var al struct {
+			Adapters []struct {
+				Name    string `json:"name"`
+				Version string `json:"version"`
+			} `json:"adapters"`
+		}
+		if err := json.Unmarshal([]byte(body), &al); err != nil {
+			t.Fatal(err)
+		}
+		for _, a := range al.Adapters {
+			if a.Name == "openai" {
+				if a.Version != want {
+					t.Fatalf("%s: openai 版本应 %q, 实得 %q", msg, want, a.Version)
+				}
+				return
+			}
+		}
+		t.Fatalf("%s: openai 适配器不存在: %s", msg, body)
+	}
+	assertOpenAIVersion("2.0.0", "现任内置必须继续服务（不被静默覆盖）")
+
+	// ④ 裁决 candidate → 插件版生效
+	arbit, _ := json.Marshal(map[string]string{
+		"kind": "adapter", "name": "openai", "winner": "candidate",
+	})
+	if resp, body = g.do(http.MethodPost, "/api/plugins/conflicts", string(arbit), true); resp.StatusCode != 200 {
+		t.Fatalf("裁决 candidate: %d %s", resp.StatusCode, body)
+	}
+
+	// ⑤ 适配器版本已切换 + 冲突清空
+	assertOpenAIVersion("SHADOW_V1", "裁决后插件版应生效")
+	resp, body = g.do(http.MethodGet, "/api/plugins/conflicts", "", true)
+	if resp.StatusCode != 200 {
+		t.Fatal(resp.StatusCode)
+	}
+	if strings.Contains(body, `"name":"openai"`) && strings.Contains(body, `"adapters":[{`) {
+		t.Fatalf("裁决后冲突应清除: %s", body)
+	}
+}

e2e/e2e_test.go

@@ -1,4 +1,4 @@
-// Package e2e runs the real llmsproxy binary against mocked upstreams over
+// Package e2e runs the real modelrouter binary against mocked upstreams over
 // real HTTP: startup, auth, chat, streaming SSE, image generation, and
 // AUTO failover when a higher-priority source fails.
 package e2e
@@ -100,15 +100,15 @@ type gatewayUnderTest struct {
 func buildBinary(t *testing.T) string {
 	t.Helper()
 	dir := t.TempDir()
-	bin := filepath.Join(dir, "llmsproxy")
+	bin := filepath.Join(dir, "modelrouter")
 	if runtime.GOOS == "windows" {
 		bin += ".exe" // go build writes the exact name; Windows needs the suffix to exec
 	}
-	out, err := exec.Command("go", "build", "-tags", "luajit", "-o", bin, "llmsproxy/cmd/llmsproxy").CombinedOutput()
+	out, err := exec.Command("go", "build", "-tags", "luajit", "-o", bin, "modelrouter/cmd/modelrouter").CombinedOutput()
 	if err != nil && runtime.GOOS == "windows" {
 		// local Windows dev box may lack LuaJIT: fall back to the bundled
 		// Lua runtime (the e2e adapters used here are passthrough-only)
-		out, err = exec.Command("go", "build", "-o", bin, "llmsproxy/cmd/llmsproxy").CombinedOutput()
+		out, err = exec.Command("go", "build", "-o", bin, "modelrouter/cmd/modelrouter").CombinedOutput()
 	}
 	if err != nil {
 		t.Fatalf("build: %v\n%s", err, out)
@@ -187,18 +187,20 @@ func writeConfig(t *testing.T, dir, listenAddr string, upstreams map[string]*moc
 	sb.WriteString("runtime_file: " + filepath.Join(dir, "runtime.json") + "\n")
 	sb.WriteString("sources:\n")
 	order := []string{"good", "fallback", "image"}
-	prio := map[string]int{"good": 100, "fallback": 50}
 	for _, name := range order {
 		u, ok := upstreams[name]
 		if !ok {
 			continue
 		}
 		if name == "image" {
-			sb.WriteString("  - name: imagegen\n    base_url: " + u.baseURL + "\n    adapter: openai\n    models:\n      - id: flux-1\n        kind: image\n        priority: 80\n")
+			sb.WriteString("  - name: imagegen\n    base_url: " + u.baseURL + "\n    adapter: openai\n    models:\n      - id: flux-1\n        kind: image\n")
 		} else {
-			sb.WriteString("  - name: " + name + "\n    base_url: " + u.baseURL + "\n    adapter: openai\n    models:\n      - id: " + name + "-m\n        priority: " + fmt.Sprint(prio[name]) + "\n")
+			sb.WriteString("  - name: " + name + "\n    base_url: " + u.baseURL + "\n    adapter: openai\n    models:\n      - id: " + name + "-m\n")
 		}
 	}
+	// AUTO 链现在是显式配置（25970a2 移除了 models.priority）。tier 越小越先试。
+	sb.WriteString("auto:\n  - model: good-m\n    source: good\n    tier: 1\n")
+	sb.WriteString("  - model: fallback-m\n    source: fallback\n    tier: 2\n")
 	path := filepath.Join(dir, "config.yaml")
 	if err := os.WriteFile(path, []byte(sb.String()), 0o644); err != nil {
 		t.Fatalf("write config: %v", err)

e2e/wire_golden_test.go

@@ -0,0 +1,192 @@
+package e2e
+
+import (
+	"bufio"
+	"encoding/json"
+	"io"
+	"net"
+	"net/http"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"testing"
+)
+
+// mustAddr 选一个空闲端口并立即释放（与 e2e_test.go 的做法一致）。
+func mustAddr(t *testing.T) string {
+	t.Helper()
+	l, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("pick port: %v", err)
+	}
+	addr := l.Addr().String()
+	l.Close()
+	return addr
+}
+
+// 本文件建立**字节级**对外输出基线，供统一线格式 v2（对外序列化移交适配器）
+// 做回归比对。
+//
+// 为什么必须有它：Phase 2 要把 gateway/chat.go 里硬编码的 OpenAI wire 结构与
+// writeChatCompletion/pumpStream 的 JSON 拼装移交给适配器 serializer。现有 e2e
+// 只用 strings.Contains 做宽松断言（"包含 Hel"），无法证明"输出字节未变"——
+// 字段集、omitempty 裁剪、usage 零值处理都可能静默漂移，而客户端会因此看到
+// 不同的 wire 形状。golden 把当前输出钉死，改动前后逐字节比对。
+//
+// 更新方式：GOLDEN_UPDATE=1 go test -tags luajit ./e2e/ -run TestWireGolden
+// 更新后必须人工审阅 diff —— golden 是契约，不是快照垃圾桶。
+
+// volatile 命中每次请求都变的字段，比对前替换为占位符。
+// 允许冒号后的空白：golden 是 pretty-print 过的（`"created": 123`）。
+var volatile = []*regexp.Regexp{
+	regexp.MustCompile(`"id":\s*"[^"]*"`),
+	regexp.MustCompile(`"created":\s*\d+`),
+}
+
+func scrub(s string) string {
+	s = volatile[0].ReplaceAllString(s, `"id": "<ID>"`)
+	return volatile[1].ReplaceAllString(s, `"created": <TS>`)
+}
+
+func compareGolden(t *testing.T, name, got string) {
+	t.Helper()
+	got = scrub(got)
+	path := filepath.Join("..", ".snapshots", name)
+	if os.Getenv("GOLDEN_UPDATE") == "1" {
+		if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(path, []byte(got), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		t.Logf("golden updated: %s", path)
+		return
+	}
+	want, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("golden missing (%s): %v\nrun GOLDEN_UPDATE=1 go test -tags luajit ./e2e/ -run TestWireGolden to create", path, err)
+	}
+	if got != string(want) {
+		t.Errorf("对外输出与 golden 漂移 (%s)\n--- want ---\n%s\n--- got ---\n%s",
+			name, string(want), got)
+	}
+}
+
+// prettyJSON 缩进输出。走 map 会排序键，从而让「字段集」成为唯一被比较的东西，
+// 忽略 Go struct 声明顺序（那不是对外契约的一部分）。
+func prettyJSON(t *testing.T, raw string) string {
+	t.Helper()
+	var v interface{}
+	if err := json.Unmarshal([]byte(raw), &v); err != nil {
+		return raw
+	}
+	out, err := json.MarshalIndent(v, "", "  ")
+	if err != nil {
+		return raw
+	}
+	return string(out)
+}
+
+// TestWireGoldenNonStreamChat 钉死非流式 /v1/chat/completions 的字段集。
+func TestWireGoldenNonStreamChat(t *testing.T) {
+	up := newMockUpstream(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	g := startGateway(t, buildBinary(t), addr, writeConfig(t, dir, addr, map[string]*mockUpstream{"good": up}))
+
+	resp, body := g.do("POST", "/v1/chat/completions",
+		`{"model":"good/good-m","messages":[{"role":"user","content":"hi"}]}`, true)
+
+	// 语义：OpenAI 客户端依赖这些顶层字段
+	var obj map[string]interface{}
+	if err := json.Unmarshal([]byte(body), &obj); err != nil {
+		t.Fatalf("not JSON: %v (%s)", err, body)
+	}
+	for _, k := range []string{"id", "object", "created", "model", "choices", "usage"} {
+		if _, ok := obj[k]; !ok {
+			t.Errorf("非流式响应缺少 OpenAI 必需字段 %q: %s", k, body)
+		}
+	}
+	if obj["object"] != "chat.completion" {
+		t.Errorf("object = %v, want chat.completion", obj["object"])
+	}
+	if resp.StatusCode != 200 {
+		t.Errorf("status = %d, want 200", resp.StatusCode)
+	}
+
+	compareGolden(t, "chat_nonstream.json", prettyJSON(t, body))
+}
+
+// TestWireGoldenStreamChat 钉死流式 SSE 的帧形状与终止序列。
+func TestWireGoldenStreamChat(t *testing.T) {
+	up := newMockUpstream(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	g := startGateway(t, buildBinary(t), addr, writeConfig(t, dir, addr, map[string]*mockUpstream{"good": up}))
+
+	resp, _ := doStream(t, g, `{"model":"good/good-m","stream":true,"messages":[{"role":"user","content":"hi"}]}`)
+	defer resp.Body.Close()
+
+	var frames []string
+	sc := bufio.NewScanner(resp.Body)
+	for sc.Scan() {
+		line := sc.Text()
+		if !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+		payload := strings.TrimPrefix(line, "data: ")
+		if payload == "[DONE]" {
+			frames = append(frames, "[DONE]")
+			continue
+		}
+		frames = append(frames, prettyJSON(t, payload))
+	}
+	if len(frames) < 2 {
+		t.Fatalf("期望多帧 SSE，实得 %d: %v", len(frames), frames)
+	}
+	if frames[len(frames)-1] != "[DONE]" {
+		t.Errorf("末帧 = %q, want [DONE]", frames[len(frames)-1])
+	}
+	// 首帧必须是 role 前导：OpenAI 客户端靠它建立 assistant 消息
+	if !strings.Contains(frames[0], `"role"`) {
+		t.Errorf("首个 SSE 帧必须携带 role 前导，实得: %s", frames[0])
+	}
+
+	compareGolden(t, "chat_stream.txt", strings.Join(frames, "\n"))
+}
+
+// TestWireGoldenImage 钉死 /v1/images/generations 的响应形状。
+func TestWireGoldenImage(t *testing.T) {
+	up := newMockUpstream(t)
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	g := startGateway(t, buildBinary(t), addr, writeConfig(t, dir, addr, map[string]*mockUpstream{"image": up}))
+
+	resp, body := g.do("POST", "/v1/images/generations", `{"model":"imagegen/flux-1","prompt":"a cat"}`, true)
+	if resp.StatusCode != 200 {
+		t.Fatalf("image status = %d: %s", resp.StatusCode, body)
+	}
+	if !strings.Contains(body, "QUJD") {
+		t.Fatalf("生图响应丢失了载荷: %s", body)
+	}
+	compareGolden(t, "image_generation.json", prettyJSON(t, body))
+}
+
+// doStream 发流式请求并返回未读完的响应（调用方负责逐行读 SSE）。
+func doStream(t *testing.T, g *gatewayUnderTest, body string) (*http.Response, error) {
+	t.Helper()
+	req, err := http.NewRequest("POST", "http://"+g.addr+"/v1/chat/completions", strings.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+g.key)
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatalf("stream request: %v", err)
+	}
+	return resp, nil
+}
+
+var _ = io.Discard

go.mod

@@ -1,4 +1,4 @@
-module llmsproxy
+module modelrouter
 
 go 1.25.0
 

internal/billing/compile.go

@@ -1,323 +0,0 @@
-package billing
-
-import (
-	"fmt"
-	"strings"
-
-	"llmsproxy/internal/config"
-)
-
-// Compile turns one billing profile into the prices table the billing plugin
-// expects, resolving URL rules against the gateway's actual sources.
-//
-// WHY URL RULES NEED RESOLVING AT ALL: the operator declares pricing by URL
-// because that is what a provider's price list is keyed on, and because several
-// sources can point at the same URL. The plugin, however, looks up by
-// `payload.source` (a source NAME) and by model — it has no idea what URL a
-// request went to. So the URL match happens here, at compile time, where the
-// config's name->base_url mapping is known, and the result is expressed in the
-// dimensions the plugin already supports.
-//
-// Profiles are the "let the user choose" axis: the same URL can appear in
-// several profiles and switching recomputes this table, so a gateway can be
-// repriced without editing the plugin.
-func Compile(profile *config.BillingProfile, sources []config.Source) (map[string]interface{}, error) {
-	return CompileOpts(profile, sources, false)
-}
-
-// CompileOpts is Compile with an explicit policy for rules that match no
-// source.
-//
-// The default is STRICT and that is correct for startup: a profile with a
-// typo'd URL prices nothing, every request on it is recorded as unpriced, and
-// the bill silently comes out wrong. Failing the load is the right response.
-//
-// The rules editor needs the opposite. An operator editing a half-finished
-// profile — adding a rule before the source that will use it exists, or
-// renaming a URL while typing — must be able to save and then see the warning,
-// not be blocked by an error they can only resolve by guessing. Lenient mode
-// keeps the offending rule in the config and drops it from the compiled table,
-// so it is preserved as text and clearly reported, without pretending it
-// prices anything.
-func CompileOpts(profile *config.BillingProfile, sources []config.Source, lenient bool) (map[string]interface{}, error) {
-	if profile == nil {
-		return nil, fmt.Errorf("no billing profile")
-	}
-	prices := map[string]interface{}{
-		"currency": "USD",
-		"default":  map[string]interface{}{"prompt": 0.0, "completion": 0.0, "per_request": 0.0},
-		"sources":  map[string]interface{}{},
-		"models":   map[string]interface{}{},
-		"keys":     map[string]interface{}{},
-	}
-	if profile.Default != "" {
-		prices["default_mode"] = profile.Default
-	}
-
-	// Currency is a profile-level statement; a rule may override it.
-	for i := range profile.Rules {
-		if c := profile.Rules[i].Currency; c != "" {
-			prices["currency"] = c
-			break
-		}
-	}
-
-	matched := map[string]bool{}
-	for i := range profile.Rules {
-		rule := &profile.Rules[i]
-		targets := matchSources(rule.URL, sources)
-		if len(targets) == 0 && rule.URL != "*" {
-			// A rule for a URL no source uses is almost always a typo (or a
-			// source that was removed). Failing loudly beats a profile that
-			// silently prices nothing — the whole reason this is a config file
-			// instead of a hand-written JSON blob.
-			if lenient {
-				continue // kept in the config text, reported as a warning
-			}
-			return nil, fmt.Errorf("rule url %q matches no configured source base_url", rule.URL)
-		}
-		for _, srcName := range targets {
-			if matched[srcName] {
-				// First rule wins. Two rules matching one source is ambiguous,
-				// and silently letting the later one win makes the file's
-				// meaning depend on ordering the operator cannot see.
-				continue
-			}
-			matched[srcName] = true
-			if err := applyRule(prices, rule, srcName, sources); err != nil {
-				return nil, err
-			}
-		}
-	}
-
-	// Sources no rule matched fall through to the profile default, which the
-	// plugin applies via prices.default. Recording them explicitly means the
-	// UI can say "this source is unpriced" instead of leaving the operator to
-	// infer it from a zero.
-	var unmatched []string
-	for _, src := range sources {
-		if !matched[src.Name] {
-			unmatched = append(unmatched, src.Name)
-		}
-	}
-	if len(unmatched) > 0 {
-		prices["unmatched_sources"] = unmatched
-	}
-	return prices, nil
-}
-
-// matchSources returns the source names whose base_url matches pattern.
-// "*" matches every source (used as a catch-all default rule).
-func matchSources(pattern string, sources []config.Source) []string {
-	var out []string
-	if pattern == "*" {
-		for _, s := range sources {
-			out = append(out, s.Name)
-		}
-		return out
-	}
-	want := normalizeURL(pattern)
-	for _, s := range sources {
-		if normalizeURL(s.BaseURL) == want {
-			out = append(out, s.Name)
-		}
-	}
-	return out
-}
-
-// normalizeURL compares URLs the way an operator expects: trailing slashes and
-// case in the host are differences the provider's price list does not care
-// about, and requiring an exact byte match would make the config brittle.
-func normalizeURL(u string) string {
-	u = strings.TrimSpace(u)
-	u = strings.TrimRight(u, "/")
-	return strings.ToLower(u)
-}
-
-// applyRule writes one rule's pricing for one source into the prices table.
-func applyRule(prices map[string]interface{}, rule *config.BillingRule, srcName string, sources []config.Source) error {
-	switch rule.Mode {
-	case "free":
-		// Explicitly priced at zero. This is NOT the same as unpriced: a source
-		// the operator says is free must not appear in the unpriced warnings,
-		// or those warnings become noise and stop being read.
-		prices["sources"].(map[string]interface{})[srcName] = map[string]interface{}{
-			"prompt": 0.0, "completion": 0.0, "per_request": 0.0,
-		}
-		return nil
-
-	case "subscription":
-		// A fixed monthly commitment: the per-request MARGINAL cost is zero, and
-		// the flat fee is reported separately. Spreading a monthly fee across
-		// requests would invent a per-request number the provider never charges,
-		// and it would change every time traffic did.
-		prices["sources"].(map[string]interface{})[srcName] = map[string]interface{}{
-			"prompt": 0.0, "completion": 0.0, "per_request": 0.0,
-		}
-		fixed, _ := prices["subscriptions"].(map[string]interface{})
-		if fixed == nil {
-			fixed = map[string]interface{}{}
-			prices["subscriptions"] = fixed
-		}
-		cur := "USD"
-		if rule.Currency != "" {
-			cur = rule.Currency
-		}
-		fixed[srcName] = map[string]interface{}{
-			"monthly": rule.Subscription, "currency": cur,
-		}
-		return nil
-
-	case "unpriced":
-		// Deliberately left out of `sources` so the plugin's unpriced_models /
-		// unpriced_reqs counters catch it. That is the point: a subscription
-		// plan whose credits cannot be converted to tokens must be VISIBLE as
-		// unbilled, not quietly estimated.
-		return nil
-
-	case "token":
-		if len(rule.Models) == 0 {
-			return fmt.Errorf("rule for url %q: mode token requires at least one model", rule.URL)
-		}
-		models := prices["models"].(map[string]interface{})
-		for modelID, t := range rule.Models {
-			entry := map[string]interface{}{}
-			prompt, err := usdPerM(t.Prompt)
-			if err != nil {
-				return fmt.Errorf("model %q prompt: %w", modelID, err)
-			}
-			completion, err := usdPerM(t.Completion)
-			if err != nil {
-				return fmt.Errorf("model %q completion: %w", modelID, err)
-			}
-			entry["prompt"] = prompt
-			entry["completion"] = completion
-			if t.CacheDiscount != nil {
-				entry["cache_discount"] = *t.CacheDiscount
-			}
-			// reasoning is optional and stays ABSENT (not nil, not 0) when the
-			// rule does not set it, which is what makes the billing plugin skip
-			// the repricing branch entirely. Writing an explicit 0 here would
-			// mean "thinking is free" and silently under-bill every request
-			// from a provider that bills it like ordinary output.
-			if strings.TrimSpace(t.Reasoning) != "" {
-				rp, err := usdPerM(t.Reasoning)
-				if err != nil {
-					return fmt.Errorf("model %q reasoning: %w", modelID, err)
-				}
-				entry["reasoning_price"] = rp
-			}
-			if lc := t.LongContext; lc != nil {
-				if lc.Above <= 0 {
-					return fmt.Errorf("model %q long_context: above must be > 0", modelID)
-				}
-				above, err := longContextRate(lc, prompt)
-				if err != nil {
-					return fmt.Errorf("model %q long_context: %w", modelID, err)
-				}
-				entry["long_context"] = map[string]interface{}{
-					"above": lc.Above,
-					// Marginal, not blended. The plugin charges the tokens past the
-					// threshold at this rate and the rest at the base rate, so the
-					// threshold composes with a cache discount (which applies to the
-					// cached portion) instead of forcing the plugin to guess how
-					// the two interact.
-					"prompt": above,
-				}
-			}
-			if rule.Peak != nil {
-				entry["peak"] = compilePeak(rule.Peak)
-			}
-			models[modelID] = entry
-		}
-		// A specific-URL token rule needs a source entry so a model served
-		// from THIS url but absent from `rule.Models` is still recognised as
-		// belonging to a priced provider.
-		//
-		// A WILDCARD rule ("*") must NOT get one. "*" is a model catalogue — it
-		// says "these model ids have these prices, wherever they come from" —
-		// and marking every source priced at {0,0} would make priceFor() set
-		// priced=true for every model the catalogue does NOT list. Paid
-		// sources (commandcode, alittokenplan, …) would then record $0 instead
-		// of showing up in unpriced_reqs, which is precisely the silent
-		// zero-billing failure this whole plugin is built to make visible.
-		// The model lookup p.models[payload.model] already marks a listed
-		// model priced on its own, so the wildcard loses nothing.
-		if rule.URL != "*" {
-			if _, ok := prices["sources"].(map[string]interface{})[srcName]; !ok {
-				prices["sources"].(map[string]interface{})[srcName] = map[string]interface{}{
-					"prompt": 0.0, "completion": 0.0, "per_request": 0.0,
-				}
-			}
-		}
-		return nil
-	}
-	return fmt.Errorf("unknown mode %q", rule.Mode)
-}
-
-// usdPerM converts a USD-per-million string into the per-token rate the plugin
-// expects. Delegates to config.ParseFloatUSDPerM so the DSL validator and the
-// compiler agree byte-for-byte on what a valid price is — a value that passes
-// Validate() but fails here (or vice versa) would be the worst kind of drift.
-func usdPerM(s string) (float64, error) {
-	return config.ParseFloatUSDPerM(s)
-}
-
-// compilePeak renders the peak window in the shape the plugin reads:
-// { multiplier, windows = { { days = {...}, hours = { {lo,hi}, ... } } } }.
-//
-// The plugin reads `days` / `hours` pairs. The first version of the deployed
-// price table used { start, end, weekdays } — a shape nothing reads — so peak
-// traffic was billed at off-peak rates with no error anywhere. Compiling from
-// typed config fields removes the chance of writing the wrong key names by hand.
-func compilePeak(p *config.BillingPeak) map[string]interface{} {
-	win := map[string]interface{}{}
-	if len(p.Weekdays) > 0 {
-		days := make([]interface{}, 0, len(p.Weekdays))
-		for _, d := range p.Weekdays {
-			days = append(days, d)
-		}
-		win["days"] = days
-	}
-	if len(p.Hours) > 0 {
-		hours := make([]interface{}, 0, len(p.Hours))
-		for _, h := range p.Hours {
-			hours = append(hours, []interface{}{h[0], h[1]})
-		}
-		win["hours"] = hours
-	}
-	out := map[string]interface{}{"multiplier": p.Multiplier}
-	if len(win) > 0 {
-		out["windows"] = []interface{}{win}
-	}
-	return out
-}
-
-// longContextRate resolves the marginal per-token rate for the tokens ABOVE a
-// threshold, given the model's base per-million prompt price.
-//
-// Two ways to express it are supported because providers publish both: an
-// absolute rate (OpenAI quotes \$3.00/M past 200k) and a multiple (DeepSeek
-// quotes exactly 2x past 128k). Neither may be missing: a step with no rate is a
-// silent no-op, and an operator who wrote one meant something.
-func longContextRate(lc *config.BillingLongContext, basePerToken float64) (float64, error) {
-	hasPrompt := strings.TrimSpace(lc.Prompt) != ""
-	hasMult := lc.Multiplier > 0
-	switch {
-	case hasPrompt && hasMult:
-		p, err := usdPerM(lc.Prompt)
-		if err != nil {
-			return 0, err
-		}
-		// Both given: the absolute rate wins, because it is the one the provider
-		// actually quotes. Silently multiplying instead would produce a number
-		// the operator never wrote.
-		return p, nil
-	case hasPrompt:
-		return usdPerM(lc.Prompt)
-	case hasMult:
-		return basePerToken * lc.Multiplier, nil
-	default:
-		return 0, fmt.Errorf("needs either prompt or multiplier")
-	}
-}

internal/billing/wildcard_test.go

@@ -1,214 +0,0 @@
-package billing
-
-import (
-	"testing"
-
-	"llmsproxy/internal/config"
-)
-
-// The wildcard token rule must not blanket-mark every source as priced.
-//
-// Why this is a hazard and not a style question: priceFor() sets priced=true as
-// soon as the request's SOURCE has an entry in the price table, and that flag
-// is what keeps a request OUT of unpriced_reqs. A "*" rule is a model
-// catalogue — it prices specific model ids wherever they come from — but
-// compiling it into a {0,0,0} entry on all 21 sources made every model the
-// catalogue does not list look "priced at zero". Paid providers (commandcode,
-// alittokenplan, …) then stopped appearing in the unpriced warnings, which is
-// exactly the silent zero-billing failure the plugin exists to surface.
-func TestWildcardTokenRuleDoesNotMarkEverySourcePriced(t *testing.T) {
-	sources := []config.Source{
-		{Name: "free1", BaseURL: "https://free1.example.com/v1"},
-		{Name: "paid1", BaseURL: "https://paid1.example.com/v1"},
-		{Name: "paid2", BaseURL: "https://paid2.example.com/v1"},
-	}
-	profile := &config.BillingProfile{
-		ID: "p",
-		Rules: []config.BillingRule{
-			{URL: "https://free1.example.com/v1", Mode: "free"},
-			{
-				URL: "*", Mode: "token",
-				Models: map[string]config.BillingToken{
-					"only-catalogued-model": {Prompt: "1.00", Completion: "2.00"},
-				},
-			},
-		},
-	}
-	prices, err := Compile(profile, sources)
-	if err != nil {
-		t.Fatalf("compile: %v", err)
-	}
-	srcs := prices["sources"].(map[string]interface{})
-
-	// The explicitly free source keeps its entry (free is a deliberate 0).
-	if _, ok := srcs["free1"]; !ok {
-		t.Error("the explicit free rule lost its source entry")
-	}
-	// Paid sources must NOT be marked, or their unlisted models bill $0.
-	for _, name := range []string{"paid1", "paid2"} {
-		if _, ok := srcs[name]; ok {
-			t.Errorf("★ wildcard token rule marked %q as priced at 0 — its unlisted "+
-				"models will bill $0 instead of appearing as unpriced", name)
-		}
-	}
-	// The catalogue itself must still be present, or the rule prices nothing.
-	models := prices["models"].(map[string]interface{})
-	if _, ok := models["only-catalogued-model"]; !ok {
-		t.Error("the wildcard rule did not publish its model price")
-	}
-}
-
-// The inverse guard: a token rule for a SPECIFIC url still marks that source,
-// because it genuinely declares pricing for the models served from it.
-func TestSpecificURLTokenRuleStillMarksItsSource(t *testing.T) {
-	sources := []config.Source{
-		{Name: "mine", BaseURL: "https://mine.example.com/v1"},
-		{Name: "other", BaseURL: "https://other.example.com/v1"},
-	}
-	profile := &config.BillingProfile{
-		ID: "p",
-		Rules: []config.BillingRule{{
-			URL: "https://mine.example.com/v1", Mode: "token",
-			Models: map[string]config.BillingToken{
-				"m": {Prompt: "1.00", Completion: "2.00"},
-			},
-		}},
-	}
-	prices, err := Compile(profile, sources)
-	if err != nil {
-		t.Fatalf("compile: %v", err)
-	}
-	srcs := prices["sources"].(map[string]interface{})
-	if _, ok := srcs["mine"]; !ok {
-		t.Error("a specific-url token rule must still mark its own source priced")
-	}
-	if _, ok := srcs["other"]; ok {
-		t.Error("an unrelated source was marked priced by a specific-url rule")
-	}
-}
-
-// TestReasoningPriceIsOmittedUnlessSet pins the absence convention for the
-// thinking-token rate.
-//
-// A rule that says nothing about reasoning must produce a price entry with NO
-// reasoning_price key, so the billing plugin skips the repricing branch and
-// bills thinking at the ordinary output rate — which is what every provider
-// actually charges. Writing an explicit 0 instead would read as "thinking is
-// free" and silently under-bill every request from that model.
-func TestReasoningPriceIsOmittedUnlessSet(t *testing.T) {
-	sources := []config.Source{{Name: "up", BaseURL: "https://up.example/v1"}}
-
-	// Case 1: no reasoning declared → the key must not appear.
-	p1 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
-		URL: "https://up.example/v1", Mode: "token",
-		Models: map[string]config.BillingToken{
-			"m": {Prompt: "1.00", Completion: "2.00"},
-		},
-	}}}
-	prices, err := Compile(p1, sources)
-	if err != nil {
-		t.Fatal(err)
-	}
-	entry := prices["models"].(map[string]interface{})["m"].(map[string]interface{})
-	if _, present := entry["reasoning_price"]; present {
-		t.Errorf("reasoning_price present (%v) without a rule; the plugin would treat 0 as free",
-			entry["reasoning_price"])
-	}
-
-	// Case 2: declared → present, and converted to per-token like every other rate.
-	p2 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
-		URL: "https://up.example/v1", Mode: "token",
-		Models: map[string]config.BillingToken{
-			"m": {Prompt: "1.00", Completion: "2.00", Reasoning: "6.00"},
-		},
-	}}}
-	prices2, err := Compile(p2, sources)
-	if err != nil {
-		t.Fatal(err)
-	}
-	e2 := prices2["models"].(map[string]interface{})["m"].(map[string]interface{})
-	rp, ok := e2["reasoning_price"].(float64)
-	if !ok {
-		t.Fatalf("reasoning_price missing or wrong type: %#v", e2["reasoning_price"])
-	}
-	if rp != 6e-6 {
-		t.Errorf("reasoning_price = %v, want 6e-6 (USD per TOKEN)", rp)
-	}
-	// A malformed rate must be rejected, not silently ignored: an operator who
-	// typed "6 USD" expects either that value or an error, never a rule that
-	// looks applied and bills nothing.
-	p3 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
-		URL: "https://up.example/v1", Mode: "token",
-		Models: map[string]config.BillingToken{
-			"m": {Prompt: "1.00", Completion: "2.00", Reasoning: "six dollars"},
-		},
-	}}}
-	if _, err := Compile(p3, sources); err == nil {
-		t.Error("a non-numeric reasoning rate must fail the compile")
-	}
-}
-
-// TestLongContextStepCompiles pins the long-context step's two accepted forms
-// and its rejection of a step that would silently do nothing.
-//
-// Providers publish both shapes — DeepSeek says "2x past 128k", OpenAI quotes
-// an absolute "$3.00/M past 200k" — so both must compile. Neither may be
-// missing: a step with no rate is a rule that LOOKS like it handles long
-// context and does not, which under-bills without any visible sign.
-func TestLongContextStepCompiles(t *testing.T) {
-	sources := []config.Source{{Name: "up", BaseURL: "https://up.example/v1"}}
-
-	compile := func(lc config.BillingLongContext) (map[string]interface{}, error) {
-		p := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
-			URL: "https://up.example/v1", Mode: "token",
-			Models: map[string]config.BillingToken{
-				"m": {Prompt: "1.00", Completion: "2.00", LongContext: &lc},
-			},
-		}}}
-		prices, err := Compile(p, sources)
-		if err != nil {
-			return nil, err
-		}
-		e := prices["models"].(map[string]interface{})["m"].(map[string]interface{})
-		return e["long_context"].(map[string]interface{}), nil
-	}
-
-	// Absolute rate (OpenAI's published shape).
-	lc1, err := compile(config.BillingLongContext{Above: 200000, Prompt: "3.00"})
-	if err != nil {
-		t.Fatalf("absolute long_context rate rejected: %v", err)
-	}
-	if lc1["above"] != 200000 || lc1["prompt"] != 3e-6 {
-		t.Errorf("absolute step = %#v, want above=200000 prompt=3e-6", lc1)
-	}
-
-	// Multiplier (DeepSeek's shape): 2x of the 1.00/M base.
-	lc2, err := compile(config.BillingLongContext{Above: 128000, Multiplier: 2})
-	if err != nil {
-		t.Fatalf("multiplier long_context rejected: %v", err)
-	}
-	if lc2["prompt"] != 2e-6 {
-		t.Errorf("multiplier step = %#v, want prompt=2e-6", lc2)
-	}
-
-	// Both given: the absolute rate wins, because that is the number the
-	// provider quotes. Multiplying instead would produce a figure the operator
-	// never wrote.
-	lc3, err := compile(config.BillingLongContext{Above: 128000, Prompt: "5.00", Multiplier: 2})
-	if err != nil {
-		t.Fatal(err)
-	}
-	if lc3["prompt"] != 5e-6 {
-		t.Errorf("both given = %#v, want the quoted absolute rate 5e-6", lc3)
-	}
-
-	// Neither: a no-op step must fail, not compile into a rule that silently
-	// bills long context at the base rate.
-	if _, err := compile(config.BillingLongContext{Above: 128000}); err == nil {
-		t.Error("a step with neither prompt nor multiplier must be rejected")
-	}
-	// A zero threshold would make every request "long".
-	if _, err := compile(config.BillingLongContext{Above: 0, Multiplier: 2}); err == nil {
-		t.Error("above=0 must be rejected")
-	}
-}

internal/config/config.go

@@ -7,7 +7,6 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
-	"strconv"
 	"strings"
 	"time"
 
@@ -30,15 +29,6 @@ type Config struct {
 	Auto          []ModelScope `yaml:"auto,omitempty"`       // AUTO 调度链规则（WebUI 优先级页编辑，chat）
 	AutoImage     []ModelScope `yaml:"auto_image,omitempty"` // AUTO 生图调度链规则（WebUI 优先级页·生图）
 	Keys          []GWKey      `yaml:"keys,omitempty"`       // 网关密钥（WebUI 密钥页管理）
-	// BillingDSL declares per-URL pricing profiles for the billing plugin.
-	//
-	// Declarative, because hand-writing the prices JSON was error-prone in ways
-	// that looked successful: the first PUT went out without the required
-	// `prices` wrapper and silently REPLACED the accumulated totals with the
-	// price table; the second used a peak-window shape the plugin does not read,
-	// so peak traffic was billed at off-peak rates — no error, the bill just
-	// quietly halved. A schema-checked config file catches both at load.
-	BillingDSL *BillingDSL `yaml:"billing,omitempty"`
 	// PluginVisibility overrides the audience a plugin declared for itself,
 	// keyed by plugin name: "admin" hides a plugin (and its state) from user
 	// keys, "all" widens one that shipped admin-only, "user,admin" names roles
@@ -62,30 +52,47 @@ const (
 	DefaultSourceConcurrency  = 8
 )
 
-// Model is a single exposed model id bound to a source, with priority used by
+// Model is a single exposed model id bound to a source.
-// AUTO auto selection (higher number = preferred).
+//
+// ★ 曾经有 `Priority int` 字段，并号称"高者优先参与 AUTO 选源"。**那是错的**：
+//
+//	AUTO 链由 `config.ModelScope.Tier` 驱动（见 scheduler.Chain），而 priority
+//	只被 `bestChatModel/bestImageModel` 用于"本源码内部该发哪个模型"，
+//	与跨源调度毫无关系。把它当成调度依据会让人误以为改个数字就能改顺序，
+//	改完发现毫无效果 —— 已彻底删除，避免再次误导。
 type Model struct {
-	ID       string                 `yaml:"id" json:"id"`
+	ID   string                 `yaml:"id" json:"id"`
-	Priority int                    `yaml:"priority" json:"priority"`
+	Kind string                 `yaml:"kind" json:"kind"` // "chat" (default) | "image"
-	Kind     string                 `yaml:"kind" json:"kind"` // "chat" (default) | "image"
+	Meta map[string]interface{} `yaml:"meta" json:"meta,omitempty"`
-	Meta     map[string]interface{} `yaml:"meta" json:"meta,omitempty"`
 }
 
 // Source describes a single upstream LLM provider.
 type Source struct {
-	Name          string            `yaml:"name" json:"name"`
+	Name      string `yaml:"name" json:"name"`
-	BaseURL       string            `yaml:"base_url" json:"base_url"`
+	BaseURL   string `yaml:"base_url" json:"base_url"`
-	APIKey        string            `yaml:"api_key" json:"api_key"`
+	APIKey    string `yaml:"api_key" json:"api_key"`
-	APIKeyEnv     string            `yaml:"api_key_env,omitempty" json:"-"` // reference to an env var holding the key (overrides api_key)
+	APIKeyEnv string `yaml:"api_key_env,omitempty" json:"-"` // reference to an env var holding the key (overrides api_key)
-	Adapter       string            `yaml:"adapter" json:"adapter"`
+	Adapter   string `yaml:"adapter" json:"adapter"`
-	Endpoint      string            `yaml:"endpoint" json:"endpoint,omitempty"`             // chat endpoint override
+	Endpoint  string `yaml:"endpoint" json:"endpoint,omitempty"` // chat endpoint override
-	ImageEndpoint string            `yaml:"image_endpoint" json:"image_endpoint,omitempty"` // image endpoint override
+	// StreamEndpoint overrides the CHAT endpoint for streaming requests. Most
-	Models        []Model           `yaml:"models" json:"models"`
+	// upstreams stream on the same path; Gemini is the exception (same path,
-	Headers       map[string]string `yaml:"headers" json:"headers,omitempty"`
+	// different verb). Declaring it here keeps that knowledge out of the kernel,
+	// which only compares opaque strings.
+	StreamEndpoint string            `yaml:"endpoint_stream" json:"endpoint_stream,omitempty"`
+	ImageEndpoint  string            `yaml:"image_endpoint" json:"image_endpoint,omitempty"` // image endpoint override
+	Models         []Model           `yaml:"models" json:"models"`
+	Headers        map[string]string `yaml:"headers" json:"headers,omitempty"`
 	// ProxyURL routes this source's HTTP(S) traffic through an explicit proxy
 	// (e.g. http://127.0.0.1:7890). Empty = direct connection. Needed for
 	// upstreams behind geo/IP blocks that only respond through a proxy.
-	ProxyURL      string                 `yaml:"proxy_url" json:"proxy_url,omitempty"`
+	ProxyURL string `yaml:"proxy_url" json:"proxy_url,omitempty"`
+	// Retry is an OPTIONAL upstream retry policy. It exists for upstreams that
+	// answer a business-level rejection with a status code the kernel would
+	// otherwise surface to the client — e.g. a per-account concurrency cap that
+	// clears by itself after tens of seconds. The kernel stays vendor-agnostic:
+	// the matcher is pure DATA (status list + response-body substrings), so no
+	// provider name ever appears in core.
+	Retry         *RetryPolicy           `yaml:"retry" json:"retry,omitempty"`
 	Meta          map[string]interface{} `yaml:"meta" json:"meta,omitempty"`
 	Temperature   float64                `yaml:"temperature" json:"temperature,omitempty"`
 	MaxTokens     int                    `yaml:"max_tokens" json:"max_tokens,omitempty"`
@@ -95,9 +102,66 @@ type Source struct {
 	RPM           int                    `yaml:"rpm" json:"rpm,omitempty"` // optional requests-per-minute cap (0 = unlimited)
 }
 
+// RetryPolicy describes when a failed upstream attempt may be retried BEFORE
+// anything is returned to the client. All conditions are data:
+//
+//	OnStatus       - retry only when the HTTP status is one of these
+//	OnBodyContains - ... AND the response body contains one of these substrings
+//	                 (empty = status alone decides)
+//	Delay          - wait between attempts ("10s" etc.); parsed on demand
+//	MaxAttempts    - TOTAL attempts including the first (1 or 0 = no retry)
+//
+// Scope: the retry happens inside one gateway request, so a stream is only
+// retried before its first chunk reaches the client (the kernel already holds
+// the first chunk back), and a non-stream call is retried as a whole.
+type RetryPolicy struct {
+	OnStatus       []int    `yaml:"on_status,omitempty" json:"on_status,omitempty"`
+	OnBodyContains []string `yaml:"on_body_contains,omitempty" json:"on_body_contains,omitempty"`
+	Delay          string   `yaml:"delay,omitempty" json:"delay,omitempty"`
+	MaxAttempts    int      `yaml:"max_attempts,omitempty" json:"max_attempts,omitempty"`
+}
+
+// ShouldRetry reports whether one failed attempt matches this policy.
+func (r *RetryPolicy) ShouldRetry(status int, body string) bool {
+	if r == nil || r.MaxAttempts <= 1 {
+		return false
+	}
+	statusOK := len(r.OnStatus) == 0
+	for _, s := range r.OnStatus {
+		if s == status {
+			statusOK = true
+			break
+		}
+	}
+	if !statusOK {
+		return false
+	}
+	if len(r.OnBodyContains) == 0 {
+		return true
+	}
+	for _, sub := range r.OnBodyContains {
+		if sub != "" && strings.Contains(body, sub) {
+			return true
+		}
+	}
+	return false
+}
+
+// DelayDuration is the configured pause between attempts (default 10s).
+func (r *RetryPolicy) DelayDuration() time.Duration {
+	if r == nil || r.Delay == "" {
+		return 10 * time.Second
+	}
+	d, err := time.ParseDuration(r.Delay)
+	if err != nil || d < 0 {
+		return 10 * time.Second
+	}
+	return d
+}
+
 // Load reads and validates a config file. When the file does not exist yet a
 // default config is generated at that path first (first-run bootstrap), so a
-// fresh binary just works: `llmsproxy -config /path/to/config.yaml`.
+// fresh binary just works: `modelrouter -config /path/to/config.yaml`.
 func Load(path string) (*Config, error) {
 	if _, err := EnsureDefault(path); err != nil {
 		return nil, err
@@ -373,13 +437,14 @@ type RuntimeConfig struct {
 // name and api_key) so the WebUI can spin up multiple key-bearing sources
 // from one shared template.
 type SourceTemplate struct {
-	Name          string            `json:"name"`
+	Name           string            `json:"name"`
-	BaseURL       string            `json:"base_url"`
+	BaseURL        string            `json:"base_url"`
-	Adapter       string            `json:"adapter"`
+	Adapter        string            `json:"adapter"`
-	Endpoint      string            `json:"endpoint,omitempty"`
+	Endpoint       string            `json:"endpoint,omitempty"`
-	ImageEndpoint string            `json:"image_endpoint,omitempty"`
+	StreamEndpoint string            `json:"endpoint_stream,omitempty"`
-	Models        []Model           `json:"models"`
+	ImageEndpoint  string            `json:"image_endpoint,omitempty"`
-	Headers       map[string]string `json:"headers,omitempty"`
+	Models         []Model           `json:"models"`
+	Headers        map[string]string `json:"headers,omitempty"`
 	// ProxyURL routes this source's HTTP(S) traffic through an explicit proxy.
 	ProxyURL      string                 `json:"proxy_url,omitempty"`
 	Meta          map[string]interface{} `json:"meta,omitempty"`
@@ -393,18 +458,18 @@ type SourceTemplate struct {
 // (full management) or "user" (sees only its own key); Models is the allowed
 // model scope with per-model token quota (0 = unlimited).
 type GWKey struct {
-	Key   string `yaml:"key" json:"key"`
+	Key  string `yaml:"key" json:"key"`
-	Role  string `yaml:"role" json:"role"`
+	Role string `yaml:"role" json:"role"`
-	Name  string `yaml:"name,omitempty" json:"name,omitempty"`
+	Name string `yaml:"name,omitempty" json:"name,omitempty"`
 	// Auto is this key's own AUTO scheduling chain. It overrides the global
 	// cfg.Auto for requests authenticated with this key; nil (or an empty
 	// list) means "inherit the global chain", so keys created before per-key
 	// chains existed keep working untouched.
-	Auto           []ModelScope `yaml:"auto,omitempty" json:"auto,omitempty"`
+	Auto      []ModelScope `yaml:"auto,omitempty" json:"auto,omitempty"`
-	Models         []ModelScope `yaml:"models,omitempty" json:"models,omitempty"`
+	Models    []ModelScope `yaml:"models,omitempty" json:"models,omitempty"`
-	Note           string       `yaml:"note,omitempty" json:"note,omitempty"`
+	Note      string       `yaml:"note,omitempty" json:"note,omitempty"`
-	CreatedAt      int64        `yaml:"created_at,omitempty" json:"created_at,omitempty"`
+	CreatedAt int64        `yaml:"created_at,omitempty" json:"created_at,omitempty"`
-	Seed           bool         `yaml:"seed,omitempty" json:"seed,omitempty"` // true if migrated from config gateway_keys
+	Seed      bool         `yaml:"seed,omitempty" json:"seed,omitempty"` // true if migrated from config gateway_keys
 }
 
 // HasOwnAuto reports whether the key declares its own AUTO chain rather than
@@ -412,182 +477,6 @@ type GWKey struct {
 // chain" so a user can clear an override back to the global default.
 func (k GWKey) HasOwnAuto() bool { return len(k.Auto) > 0 }
 
-// BillingDSL holds declarative per-URL pricing profiles for the billing plugin.
-//
-// Profiles are the "let the user pick" axis: the same upstream URL can be
-// declared under multiple profiles, and the operator chooses which is active.
-// Switching recomputes and re-injects the billing plugin's prices table.
-type BillingDSL struct {
-	// Active is the id of the profile in effect. Empty => the first profile.
-	Active   string           `yaml:"active,omitempty" json:"active,omitempty"`
-	Profiles []BillingProfile `yaml:"profiles,omitempty" json:"profiles,omitempty"`
-}
-
-// BillingProfile is one named set of per-URL pricing rules.
-type BillingProfile struct {
-	ID    string `yaml:"id" json:"id"`
-	Label string `yaml:"label,omitempty" json:"label,omitempty"`
-	// Default applies when no rule matches: "free" or "unpriced".
-	Default string        `yaml:"default,omitempty" json:"default,omitempty"`
-	Rules   []BillingRule `yaml:"rules,omitempty" json:"rules,omitempty"`
-}
-
-// BillingRule declares the pricing for everything under one URL.
-type BillingRule struct {
-	// URL matches a source's base_url. "*" = any unmatched URL.
-	URL string `yaml:"url" json:"url"`
-	// Mode is one of: free, token, subscription, unpriced.
-	Mode     string `yaml:"mode" json:"mode"`
-	Currency string `yaml:"currency,omitempty" json:"currency,omitempty"`
-	// Subscription is the flat monthly cost (for mode=subscription).
-	Subscription float64 `yaml:"subscription,omitempty" json:"subscription,omitempty"`
-	// Models declares per-token prices when mode=token.
-	Models map[string]BillingToken `yaml:"models,omitempty" json:"models,omitempty"`
-	// Free is true when mode=free (kept for explicitness).
-	// Peak applies to ALL token-priced requests under this URL.
-	Peak *BillingPeak `yaml:"peak,omitempty" json:"peak,omitempty"`
-}
-
-// BillingToken is the per-model price.
-type BillingToken struct {
-	Prompt        string   `yaml:"prompt" json:"prompt"`         // USD per million tokens, e.g. "0.15"
-	Completion    string   `yaml:"completion" json:"completion"` // USD per million tokens
-	CacheDiscount *float64 `yaml:"cache_discount,omitempty" json:"cache_discount,omitempty"`
-	// Reasoning is the USD-per-million rate for thinking tokens, for the rare
-	// provider that charges them differently from ordinary output. Omit it and
-	// reasoning is billed at Completion — which is what OpenAI, DeepSeek and
-	// Anthropic all do, and what every source on this deployment reports.
-	//
-	// It is a rate for a SUBSET of completion tokens, never an extra charge: the
-	// billing plugin applies only the difference between this and Completion, so
-	// setting it to a multiple of Completion reprices the thinking share rather
-	// than billing it twice.
-	Reasoning string `yaml:"reasoning,omitempty" json:"reasoning,omitempty"`
-	// LongContext is the prompt-price step above a token threshold. Nil = one
-	// flat rate, which is what a rule means when it says nothing.
-	LongContext *BillingLongContext `yaml:"long_context,omitempty" json:"long_context,omitempty"`
-}
-
-// BillingPeak declares peak-hour pricing.
-type BillingPeak struct {
-	Multiplier float64 `yaml:"multiplier" json:"multiplier"`
-	// Weekdays: 1=Mon..5=Fri (Lua os.date !%w, 0=Sun)
-	Weekdays []int `yaml:"weekdays,omitempty" json:"weekdays,omitempty"`
-	// Hours: each pair is [lo,hi] inclusive
-	Hours [][2]int `yaml:"hours,omitempty" json:"hours,omitempty"`
-}
-
-// BillingLongContext is a prompt-price step for long inputs.
-//
-// Every major provider charges the input tokens ABOVE a threshold at a higher
-// rate — DeepSeek doubles past 128k, OpenAI and Anthropic step up as well. A
-// flat prompt price therefore under-bills exactly the traffic an agent workload
-// generates, and it under-bills it silently: the number is plausible, the bill
-// adds up, and nothing says the rate was wrong.
-//
-// The step applies to the PROMPT only. Output and cache reads are billed at
-// their own rates regardless of input length, which is how providers publish
-// it — the higher rate is for the extra context you made the model read, not for
-// what it wrote back.
-type BillingLongContext struct {
-	// Above is the token count the step starts at. A request with exactly this
-	// many prompt tokens is still at the base rate; the step covers the tokens
-	// BEYOND it. Inclusive-below matches how the providers document it.
-	Above int `yaml:"above" json:"above"`
-	// Prompt is the USD-per-million rate for the tokens above Above.
-	Prompt string `yaml:"prompt" json:"prompt"`
-	// Multiplier is an alternative to Prompt: a factor applied to the base rate
-	// (DeepSeek's published step is exactly 2x). When both are set Prompt wins,
-	// and setting neither is a validation error rather than a silent no-op.
-	Multiplier float64 `yaml:"multiplier,omitempty" json:"multiplier,omitempty"`
-}
-
-// Validate rejects a BillingDSL that could not work as written.
-func (b *BillingDSL) Validate() error {
-	if b == nil {
-		return nil
-	}
-	ids := map[string]bool{}
-	for i := range b.Profiles {
-		p := &b.Profiles[i]
-		if p.ID == "" {
-			return fmt.Errorf("billing profile [%d] has no id", i)
-		}
-		if ids[p.ID] {
-			return fmt.Errorf("billing profile id %q appears more than once", p.ID)
-		}
-		ids[p.ID] = true
-		if p.Default != "" && p.Default != "free" && p.Default != "unpriced" {
-			return fmt.Errorf("profile %q: default must be free or unpriced, got %q", p.ID, p.Default)
-		}
-		for j := range p.Rules {
-			r := &p.Rules[j]
-			if r.URL == "" {
-				return fmt.Errorf("profile %q rule [%d]: url is required", p.ID, j)
-			}
-			switch r.Mode {
-			case "free", "token", "subscription", "unpriced":
-			default:
-				return fmt.Errorf("profile %q rule [%d]: mode must be free/token/subscription/unpriced, got %q", p.ID, j, r.Mode)
-			}
-			if r.Mode == "token" {
-				for m, t := range r.Models {
-					if t.Prompt == "" || t.Completion == "" {
-						return fmt.Errorf("profile %q model %q: prompt and completion are required for token mode", p.ID, m)
-					}
-					if _, err := ParseFloatUSDPerM(t.Prompt); err != nil {
-						return fmt.Errorf("profile %q model %q prompt: %w", p.ID, m, err)
-					}
-					if _, err := ParseFloatUSDPerM(t.Completion); err != nil {
-						return fmt.Errorf("profile %q model %q completion: %w", p.ID, m, err)
-					}
-				}
-			}
-		}
-	}
-	// Active must resolve to a profile id.
-	if b.Active != "" && !ids[b.Active] {
-		return fmt.Errorf("billing.active = %q but no profile has that id", b.Active)
-	}
-	return nil
-}
-
-// Resolve returns the profile with the given id; empty id (or an id that does
-// not exist) falls back to the FIRST profile. A config with profiles but no
-// active marker should mean "use the one I wrote first" rather than "price
-// nothing", which is what a nil here would silently do.
-func (b *BillingDSL) Resolve(id string) *BillingProfile {
-	if b == nil || len(b.Profiles) == 0 {
-		return nil
-	}
-	if id != "" {
-		for i := range b.Profiles {
-			if b.Profiles[i].ID == id {
-				return &b.Profiles[i]
-			}
-		}
-	}
-	return &b.Profiles[0]
-}
-
-// parseFloatUSDPerM parses a USD-per-million price string. Returns the
-// per-single-token rate the plugin expects (price / 1e6).
-//
-// Kept as a string in the DSL rather than a float: "0.15" survives round-trips
-// through YAML editors, and a malformed value must fail VALIDATION (with the
-// file and field named) rather than parse as 0 and under-bill silently — which
-// is exactly the failure mode that motivated the DSL.
-func ParseFloatUSDPerM(s string) (float64, error) {
-	f, err := strconv.ParseFloat(strings.TrimSpace(s), 64)
-	if err != nil {
-		return 0, fmt.Errorf("not a number: %q", s)
-	}
-	if f < 0 {
-		return 0, fmt.Errorf("negative price: %q", s)
-	}
-	return f / 1e6, nil
-}
-
 // KeyQuota is retained only to carry a scope entry's caps through the admin
 // API. Quotas are per model, never per key: there is deliberately no key-wide
 // total, so exhausting one model's budget never blocks the others.

internal/config/config_test.go

@@ -52,9 +52,6 @@ sources:
 	if cfg.Sources[1].MaxConcurrent == 0 {
 		t.Fatal("default max_concurrent not applied")
 	}
-	if cfg.Sources[0].Models[0].Priority != 100 {
-		t.Fatalf("priority = %d", cfg.Sources[0].Models[0].Priority)
-	}
 	if cfg.DefaultModel != "AUTO" {
 		t.Fatalf("default model = %q", cfg.DefaultModel)
 	}

internal/config/presets.go

@@ -14,9 +14,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "deepseek",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "deepseek-v4-flash", Priority: 100, Kind: "chat"},
+			{ID: "deepseek-v4-flash", Kind: "chat"},
-			{ID: "deepseek-v4-pro", Priority: 90, Kind: "chat"},
+			{ID: "deepseek-v4-pro", Kind: "chat"},
-			{ID: "deepseek-v4-flash-vision-exp", Priority: 60, Kind: "chat"},
+			{ID: "deepseek-v4-flash-vision-exp", Kind: "chat"},
 		},
 	},
 	{
@@ -25,10 +25,10 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "glm-5.3", Priority: 100, Kind: "chat"},
+			{ID: "glm-5.3", Kind: "chat"},
-			{ID: "glm-5.3-flash", Priority: 90, Kind: "chat"},
+			{ID: "glm-5.3-flash", Kind: "chat"},
-			{ID: "glm-5.2", Priority: 80, Kind: "chat"},
+			{ID: "glm-5.2", Kind: "chat"},
-			{ID: "glm-5.1", Priority: 60, Kind: "chat"},
+			{ID: "glm-5.1", Kind: "chat"},
 		},
 	},
 	{
@@ -37,10 +37,10 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "kimi-k3", Priority: 100, Kind: "chat"},
+			{ID: "kimi-k3", Kind: "chat"},
-			{ID: "kimi-k2.7-code", Priority: 90, Kind: "chat"},
+			{ID: "kimi-k2.7-code", Kind: "chat"},
-			{ID: "kimi-k2.7-code-highspeed", Priority: 85, Kind: "chat"},
+			{ID: "kimi-k2.7-code-highspeed", Kind: "chat"},
-			{ID: "kimi-k2.6", Priority: 70, Kind: "chat"},
+			{ID: "kimi-k2.6", Kind: "chat"},
 		},
 	},
 	{
@@ -49,11 +49,11 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "deepseek-ai/DeepSeek-V4-Pro", Priority: 100, Kind: "chat"},
+			{ID: "deepseek-ai/DeepSeek-V4-Pro", Kind: "chat"},
-			{ID: "deepseek-ai/DeepSeek-V4-Flash", Priority: 90, Kind: "chat"},
+			{ID: "deepseek-ai/DeepSeek-V4-Flash", Kind: "chat"},
-			{ID: "moonshotai/Kimi-K2.6", Priority: 80, Kind: "chat"},
+			{ID: "moonshotai/Kimi-K2.6", Kind: "chat"},
-			{ID: "zai-org/GLM-5.1", Priority: 70, Kind: "chat"},
+			{ID: "zai-org/GLM-5.1", Kind: "chat"},
-			{ID: "Kwai-Kolors/Kolors", Priority: 50, Kind: "image"},
+			{ID: "Kwai-Kolors/Kolors", Kind: "image"},
 		},
 	},
 	{
@@ -62,9 +62,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 4,
 		Models: []Model{
-			{ID: "4.0Ultra", Priority: 100, Kind: "chat"},
+			{ID: "4.0Ultra", Kind: "chat"},
-			{ID: "generalv3.5", Priority: 80, Kind: "chat"},
+			{ID: "generalv3.5", Kind: "chat"},
-			{ID: "max-32k", Priority: 70, Kind: "chat"},
+			{ID: "max-32k", Kind: "chat"},
 		},
 	},
 	{
@@ -73,11 +73,11 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "qwen3.8-max", Priority: 100, Kind: "chat"},
+			{ID: "qwen3.8-max", Kind: "chat"},
-			{ID: "qwen3.8-flash", Priority: 90, Kind: "chat"},
+			{ID: "qwen3.8-flash", Kind: "chat"},
-			{ID: "qwen3.7-plus", Priority: 85, Kind: "chat"},
+			{ID: "qwen3.7-plus", Kind: "chat"},
-			{ID: "deepseek-v4-pro", Priority: 70, Kind: "chat"},
+			{ID: "deepseek-v4-pro", Kind: "chat"},
-			{ID: "kimi-k3", Priority: 70, Kind: "chat"},
+			{ID: "kimi-k3", Kind: "chat"},
 		},
 	},
 	{
@@ -86,9 +86,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "ernie-5.0-turbo", Priority: 100, Kind: "chat"},
+			{ID: "ernie-5.0-turbo", Kind: "chat"},
-			{ID: "ernie-5.0", Priority: 90, Kind: "chat"},
+			{ID: "ernie-5.0", Kind: "chat"},
-			{ID: "deepseek-v4-pro", Priority: 70, Kind: "chat"},
+			{ID: "deepseek-v4-pro", Kind: "chat"},
 		},
 	},
 	{
@@ -97,9 +97,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "doubao-seed-2.0-pro", Priority: 100, Kind: "chat"},
+			{ID: "doubao-seed-2.0-pro", Kind: "chat"},
-			{ID: "doubao-seed-2.0", Priority: 90, Kind: "chat"},
+			{ID: "doubao-seed-2.0", Kind: "chat"},
-			{ID: "doubao-seed-2.0-flash", Priority: 80, Kind: "chat"},
+			{ID: "doubao-seed-2.0-flash", Kind: "chat"},
 		},
 	},
 	{
@@ -108,8 +108,8 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "MiniMax-M3", Priority: 100, Kind: "chat"},
+			{ID: "MiniMax-M3", Kind: "chat"},
-			{ID: "MiniMax-M2.5", Priority: 80, Kind: "chat"},
+			{ID: "MiniMax-M2.5", Kind: "chat"},
 		},
 	},
 	// ── 国际主流 ────────────────────────────────────────────────────────────
@@ -119,9 +119,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "openai",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "gpt-5.6-sol", Priority: 100, Kind: "chat"},
+			{ID: "gpt-5.6-sol", Kind: "chat"},
-			{ID: "gpt-5.6-terra", Priority: 90, Kind: "chat"},
+			{ID: "gpt-5.6-terra", Kind: "chat"},
-			{ID: "gpt-5.6-luna", Priority: 80, Kind: "chat"},
+			{ID: "gpt-5.6-luna", Kind: "chat"},
 		},
 	},
 	{
@@ -130,10 +130,10 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "anthropic",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "claude-opus-5", Priority: 100, Kind: "chat"},
+			{ID: "claude-opus-5", Kind: "chat"},
-			{ID: "claude-fable-5", Priority: 95, Kind: "chat"},
+			{ID: "claude-fable-5", Kind: "chat"},
-			{ID: "claude-sonnet-5", Priority: 85, Kind: "chat"},
+			{ID: "claude-sonnet-5", Kind: "chat"},
-			{ID: "claude-haiku-4-5", Priority: 70, Kind: "chat"},
+			{ID: "claude-haiku-4-5", Kind: "chat"},
 		},
 	},
 	{
@@ -142,10 +142,10 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "gemini",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "gemini-3.6-flash", Priority: 100, Kind: "chat"},
+			{ID: "gemini-3.6-flash", Kind: "chat"},
-			{ID: "gemini-3.5-flash", Priority: 90, Kind: "chat"},
+			{ID: "gemini-3.5-flash", Kind: "chat"},
-			{ID: "gemini-3.5-flash-lite", Priority: 80, Kind: "chat"},
+			{ID: "gemini-3.5-flash-lite", Kind: "chat"},
-			{ID: "gemini-3.1-pro", Priority: 85, Kind: "chat"},
+			{ID: "gemini-3.1-pro", Kind: "chat"},
 		},
 	},
 	{
@@ -154,9 +154,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "groq",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "openai/gpt-oss-120b", Priority: 100, Kind: "chat"},
+			{ID: "openai/gpt-oss-120b", Kind: "chat"},
-			{ID: "openai/gpt-oss-20b", Priority: 80, Kind: "chat"},
+			{ID: "openai/gpt-oss-20b", Kind: "chat"},
-			{ID: "groq/compound", Priority: 70, Kind: "chat"},
+			{ID: "groq/compound", Kind: "chat"},
 		},
 	},
 	{
@@ -165,9 +165,9 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "mistral",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "mistral-medium-latest", Priority: 100, Kind: "chat"},
+			{ID: "mistral-medium-latest", Kind: "chat"},
-			{ID: "mistral-large-latest", Priority: 90, Kind: "chat"},
+			{ID: "mistral-large-latest", Kind: "chat"},
-			{ID: "codestral-latest", Priority: 85, Kind: "chat"},
+			{ID: "codestral-latest", Kind: "chat"},
 		},
 	},
 	{
@@ -180,10 +180,10 @@ var PresetTemplates = []SourceTemplate{
 			"X-Title":      "ModelRouter",
 		},
 		Models: []Model{
-			{ID: "openrouter/auto", Priority: 100, Kind: "chat"},
+			{ID: "openrouter/auto", Kind: "chat"},
-			{ID: "deepseek/deepseek-v4-flash", Priority: 90, Kind: "chat"},
+			{ID: "deepseek/deepseek-v4-flash", Kind: "chat"},
-			{ID: "anthropic/claude-opus-5", Priority: 85, Kind: "chat"},
+			{ID: "anthropic/claude-opus-5", Kind: "chat"},
-			{ID: "openai/gpt-5.6-sol", Priority: 85, Kind: "chat"},
+			{ID: "openai/gpt-5.6-sol", Kind: "chat"},
 		},
 	},
 	{
@@ -192,8 +192,8 @@ var PresetTemplates = []SourceTemplate{
 		Adapter:       "github",
 		MaxConcurrent: 8,
 		Models: []Model{
-			{ID: "gpt-5.6-sol", Priority: 100, Kind: "chat"},
+			{ID: "gpt-5.6-sol", Kind: "chat"},
-			{ID: "gpt-5.6-luna", Priority: 80, Kind: "chat"},
+			{ID: "gpt-5.6-luna", Kind: "chat"},
 		},
 	},
 	// ── 本地模型 ────────────────────────────────────────────────────────────
@@ -204,9 +204,9 @@ var PresetTemplates = []SourceTemplate{
 		Endpoint:      "/api/chat",
 		MaxConcurrent: 4,
 		Models: []Model{
-			{ID: "qwen3", Priority: 80, Kind: "chat"},
+			{ID: "qwen3", Kind: "chat"},
-			{ID: "deepseek-r1", Priority: 80, Kind: "chat"},
+			{ID: "deepseek-r1", Kind: "chat"},
-			{ID: "llama3.3", Priority: 70, Kind: "chat"},
+			{ID: "llama3.3", Kind: "chat"},
 		},
 	},
 }

internal/config/retry_policy_test.go

@@ -0,0 +1,59 @@
+package config
+
+import (
+	"testing"
+	"time"
+)
+
+// ShouldRetry 是纯数据判据：状态码列表 + 响应体子串。它决定"这次失败值不值得
+// 在网关内部等一会儿重试"。厂商知识绝不能进内核，所以判据必须能用一个表格测清楚。
+func TestRetryPolicyShouldRetry(t *testing.T) {
+	cases := []struct {
+		name   string
+		policy *RetryPolicy
+		status int
+		body   string
+		want   bool
+	}{
+		{"nil policy", nil, 500, "anything", false},
+		{"max_attempts=1 disables", &RetryPolicy{OnStatus: []int{400}, MaxAttempts: 1}, 400, "x", false},
+		{"status matches, no body filter",
+			&RetryPolicy{OnStatus: []int{429}, MaxAttempts: 3}, 429, "", true},
+		{"status mismatch",
+			&RetryPolicy{OnStatus: []int{429}, MaxAttempts: 3}, 400, "429", false},
+		{"status + body match",
+			&RetryPolicy{OnStatus: []int{400}, OnBodyContains: []string{"TM.00001041"}, MaxAttempts: 3},
+			400, `{"error_code":"TM.00001041"}`, true},
+		{"status matches but body does not",
+			&RetryPolicy{OnStatus: []int{400}, OnBodyContains: []string{"TM.00001041"}, MaxAttempts: 3},
+			400, `{"error_code":"TM.00001005"}`, false},
+		{"empty substring never matches",
+			&RetryPolicy{OnStatus: []int{400}, OnBodyContains: []string{""}, MaxAttempts: 3}, 400, "x", false},
+		{"any of several body signals",
+			&RetryPolicy{OnStatus: []int{400, 429}, OnBodyContains: []string{"TM.00001041", "TPM limit"}, MaxAttempts: 3},
+			429, `{"message":"TPM limit exceeded"}`, true},
+	}
+	for _, tc := range cases {
+		if got := tc.policy.ShouldRetry(tc.status, tc.body); got != tc.want {
+			t.Errorf("%s: ShouldRetry(%d, %q) = %v, want %v", tc.name, tc.status, tc.body, got, tc.want)
+		}
+	}
+}
+
+func TestRetryPolicyDelay(t *testing.T) {
+	cases := []struct {
+		name   string
+		policy *RetryPolicy
+		want   time.Duration
+	}{
+		{"nil", nil, 10 * time.Second},
+		{"explicit", &RetryPolicy{Delay: "250ms"}, 250 * time.Millisecond},
+		{"garbage falls back", &RetryPolicy{Delay: "not-a-duration"}, 10 * time.Second},
+		{"negative falls back", &RetryPolicy{Delay: "-5s"}, 10 * time.Second},
+	}
+	for _, tc := range cases {
+		if got := tc.policy.DelayDuration(); got != tc.want {
+			t.Errorf("%s: DelayDuration() = %v, want %v", tc.name, got, tc.want)
+		}
+	}
+}

internal/config/secret.go

@@ -27,8 +27,9 @@ func NewGatewayKey() (string, error) {
 
 // SecretBox encrypts and decrypts sensitive values (upstream API keys,
 // custom header values, gateway keys) for persist-time protection. The
-// master key comes from the LLMS_PROXY_MASTER_KEY environment variable
+// master key comes from the MODELROUTER_MASTER_KEY environment variable
-// (hex, 64 chars) or from a generated <runtime-dir>/master.key file.
+// (hex, 64 chars; the pre-rename LLMS_PROXY_MASTER_KEY is still accepted) or
+// from a generated <runtime-dir>/master.key file.
 type SecretBox struct {
 	aead cipher.AEAD
 }
@@ -37,10 +38,18 @@ type SecretBox struct {
 // file next to the runtime file. If neither exists, a fresh random key is
 // generated and written to master.key with mode 0600.
 func NewSecretBox(runtimeFile string) (*SecretBox, error) {
-	if v := os.Getenv("LLMS_PROXY_MASTER_KEY"); v != "" {
+	// New name first, then the pre-rename name: an existing unit file that still
-		key, err := hex.DecodeString(strings.TrimSpace(v))
+	// exports LLMS_PROXY_MASTER_KEY must keep starting after an upgrade.
+	envKey := os.Getenv("MODELROUTER_MASTER_KEY")
+	envName := "MODELROUTER_MASTER_KEY"
+	if envKey == "" {
+		envKey = os.Getenv("LLMS_PROXY_MASTER_KEY")
+		envName = "LLMS_PROXY_MASTER_KEY"
+	}
+	if envKey != "" {
+		key, err := hex.DecodeString(strings.TrimSpace(envKey))
 		if err != nil || len(key) != 32 {
-			return nil, fmt.Errorf("LLMS_PROXY_MASTER_KEY: expected 64 hex chars")
+			return nil, fmt.Errorf("%s: expected 64 hex chars", envName)
 		}
 		return newBox(key)
 	}

internal/core/codearts_sig_expected.json

@@ -0,0 +1,38 @@
+[
+ {
+  "method": "GET",
+  "url": "https://snap-access.cn-north-4.myhuaweicloud.com/snap-manager/v1/statistics/plugin",
+  "body": "",
+  "ak": "AK1",
+  "sk": "SK1",
+  "st": "TOK1",
+  "canonical": "GET\n/snap-manager/v1/statistics/plugin/\n\nhost:snap-access.cn-north-4.myhuaweicloud.com\nx-sdk-content-sha256:e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855\nx-sdk-date:20261013T150405Z\nx-security-token:TOK1\n\nhost;x-sdk-content-sha256;x-sdk-date;x-security-token\ne3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+  "string_to_sign": "SDK-HMAC-SHA256\n20261013T150405Z\nf40918a3339c9db2251933dcce410db05ff8a403de2767bc642140ea93cdc182",
+  "signature": "5332d627c3b8f14973afe41905f23dce75eaf9a34134dff670cb19835beb0549",
+  "signed_headers": "host;x-sdk-content-sha256;x-sdk-date;x-security-token"
+ },
+ {
+  "method": "GET",
+  "url": "https://snap-access.cn-north-4.myhuaweicloud.com/v1/ops/delivery?channel=IDE",
+  "body": "",
+  "ak": "AK2",
+  "sk": "SK2",
+  "st": "",
+  "canonical": "GET\n/v1/ops/delivery/\nchannel=IDE\nhost:snap-access.cn-north-4.myhuaweicloud.com\nx-sdk-content-sha256:e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855\nx-sdk-date:20261013T150405Z\nx-security-token:\n\nhost;x-sdk-content-sha256;x-sdk-date;x-security-token\ne3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+  "string_to_sign": "SDK-HMAC-SHA256\n20261013T150405Z\n3c849b40bbad69b981eaf280e162878bf00687d08cb45df43ac792fc23add4f1",
+  "signature": "a01248a4121fd943f0f3f07aac101cd649ddf34d162729bf06a6a12ae9916b1a",
+  "signed_headers": "host;x-sdk-content-sha256;x-sdk-date;x-security-token"
+ },
+ {
+  "method": "POST",
+  "url": "https://snap-access.cn-north-4.myhuaweicloud.com/v1/ops/claim",
+  "body": "{\"campaignId\":\"C1\",\"channel\":\"IDE\"}",
+  "ak": "AK3",
+  "sk": "SK3",
+  "st": "T3",
+  "canonical": "POST\n/v1/ops/claim/\n\ncontent-type:application/json\nhost:snap-access.cn-north-4.myhuaweicloud.com\nx-sdk-content-sha256:85e6ce395b620ecb748af3b217f47107ebb5906d1a83fe7db221d9845de29f19\nx-sdk-date:20261013T150405Z\nx-security-token:T3\n\ncontent-type;host;x-sdk-content-sha256;x-sdk-date;x-security-token\n85e6ce395b620ecb748af3b217f47107ebb5906d1a83fe7db221d9845de29f19",
+  "string_to_sign": "SDK-HMAC-SHA256\n20261013T150405Z\n554c2607cff324a770b55e78e2a1c58afe466abf5e4eaa104a6e1fa96acfc733",
+  "signature": "e5faa52ec51c36b38804aac45a8a19876314cec465ab8f905100e657e8ea68d5",
+  "signed_headers": "content-type;host;x-sdk-content-sha256;x-sdk-date;x-security-token"
+ }
+]

internal/core/core.go

@@ -8,20 +8,18 @@ import (
 	"encoding/hex"
 	"fmt"
 	"log"
+	"net/url"
 	"os"
 	"path/filepath"
-	"sort"
-	"net/url"
 	"strings"
 	"sync"
 	"sync/atomic"
 	"time"
 
-	"llmsproxy/internal/billing"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/config"
+	"modelrouter/internal/lua"
-	"llmsproxy/internal/lua"
+	"modelrouter/internal/provider"
-	"llmsproxy/internal/provider"
+	"modelrouter/internal/scheduler"
-	"llmsproxy/internal/scheduler"
 )
 
 // Core owns the running configuration and adapters. mu guards every read and
@@ -69,6 +67,11 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	// to boot the gateway because a third-party .lua has a typo would be the
 	// wrong trade.
 	c.plugins = lua.NewPlugins(c.vm, cfg.PluginDir)
+	// R4-c: hand plugins the kernel action surface. Injected as an interface
+	// because lua must not import core (core depends on lua, not the reverse).
+	// Wired BEFORE LoadDir so a plugin that calls an action during its own
+	// __init/first hook already finds a working kernel rather than a nil one.
+	c.plugins.SetKernelAPI(c)
 	// Plugin outbound HTTP is scoped to the hosts of configured sources. Set
 	// before LoadDir so a plugin that requests on its first hook already sees
 	// the real allowlist rather than the deny-all placeholder.
@@ -92,9 +95,6 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	for name, aud := range cfg.PluginVisibility {
 		c.plugins.SetPluginVisibility(name, aud)
 	}
-	if err := c.plugins.SeedBundled(); err != nil {
-		log.Printf("[core] seed bundled plugins: %v", err)
-	}
 	if err := c.plugins.LoadDir(); err != nil {
 		log.Printf("[core] plugin dir: %v", err)
 	}
@@ -106,7 +106,6 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 		log.Printf("[core] plugin %v %v loaded (hooks=%v ui=%v)",
 			p["name"], p["version"], p["hooks"], p["ui"])
 	}
-	c.applyBillingDSL()
 	c.store = config.NewStore(cfg.RuntimeFile)
 	// Share one box between the runtime store and config.yaml so a single
 	// master.key seals both files. config.Load left the config holding
@@ -126,6 +125,11 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	// only appeared to work because seedKeys' Save() unsealed memory as a side
 	// effect. Removing the redundant saves exposed the real ordering bug.
 	cfg.AttachSecretBox(c.store.SecretBox())
+	// Share that SAME box with the plugin credential vault: one master.key seals
+	// config.yaml, the runtime store, AND every .secrets/*.json. Set here (not at
+	// NewPlugins time) because the store — and thus the box — only exists now;
+	// plugins write secrets only during login flows, well after this point.
+	c.plugins.SetSecretBox(c.store.SecretBox())
 	hadPlaintextSecrets, err := cfg.UnsealSecrets(c.store.SecretBox())
 	if err != nil {
 		return nil, fmt.Errorf("unseal secrets: %w", err)
@@ -140,9 +144,6 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	if err := c.seedKeys(); err != nil {
 		return nil, err
 	}
-	if err := c.seedAuto(); err != nil {
-		return nil, err
-	}
 	if err := c.rebuildRegistry(); err != nil {
 		return nil, err
 	}
@@ -185,39 +186,6 @@ func (c *Core) migrateFromRuntime() {
 	}
 }
 
-// seedAuto uses the existing auto rules from config.yaml, or creates an
-// initial chain from legacy per-source model priority (first run only).
-func (c *Core) seedAuto() error {
-	if len(c.cfg.Auto) > 0 {
-		return nil
-	}
-	type item struct {
-		model string
-		prio  int
-	}
-	var flat []item
-	for _, s := range c.mergedSources() {
-		for _, m := range s.Models {
-			if m.Kind == "image" {
-				continue
-			}
-			flat = append(flat, item{m.ID, m.Priority})
-		}
-	}
-	sort.SliceStable(flat, func(i, j int) bool {
-		if flat[i].prio != flat[j].prio {
-			return flat[i].prio > flat[j].prio
-		}
-		return flat[i].model < flat[j].model
-	})
-	entries := make([]config.ModelScope, 0, len(flat))
-	for _, it := range flat {
-		entries = append(entries, config.ModelScope{Model: it.model})
-	}
-	c.cfg.Auto = entries
-	return c.cfg.Save()
-}
-
 // seedKeys ensures config.yaml has admin keys for every gateway_keys entry
 // from the YAML config (once). After first save they become GWKey records
 // in c.cfg.Keys and gateway_keys is no longer read for auth.
@@ -650,11 +618,29 @@ func (c *Core) rebuildRegistry() error {
 // allowlist for plugin outbound HTTP. Unparseable URLs are skipped rather than
 // failing startup: a malformed base_url is already rejected when the provider is
 // built, and it must not take plugins down with it.
+//
+// A source may also imply a SECOND host. OpenDesign splits inference
+// (amr-link.open-design.ai) from billing (amr-api.open-design.ai), and the
+// quota plugin derives the billing host from the inference one — so the host it
+// asks for is not the host the operator configured, and the request would be
+// refused with a message that reads like a plugin defect. Adding the derived
+// host here is what makes the probe reachable.
+//
+// The grant stays scoped: only a source that IS on amr-link pulls in amr-api, so
+// a deployment with no OpenDesign source does not silently gain the entry.
 func sourceHosts(cfg *config.Config) []string {
 	if cfg == nil {
 		return nil
 	}
 	out := make([]string, 0, len(cfg.Sources))
+	seen := make(map[string]bool, len(cfg.Sources))
+	add := func(h string) {
+		if h == "" || seen[h] {
+			return
+		}
+		seen[h] = true
+		out = append(out, h)
+	}
 	for _, s := range cfg.Sources {
 		if s.BaseURL == "" {
 			continue
@@ -663,11 +649,27 @@ func sourceHosts(cfg *config.Config) []string {
 		if err != nil || u.Hostname() == "" {
 			continue
 		}
-		out = append(out, u.Hostname())
+		add(u.Hostname())
+		add(deriveBillingHost(u.Hostname()))
 	}
 	return out
 }
 
+// deriveBillingHost returns the billing host a configured source implies, or
+// the host unchanged when it implies none.
+//
+// Kept in Go as well as in the Lua adapter because the two must agree: Go builds
+// the allowlist and Lua builds the request URL, and a disagreement shows up as a
+// refused request rather than as an error naming the mismatch. The rule is the
+// full hostname, not a prefix — `amr-link.` as a prefix would also rewrite
+// amr-link.example.com into an OpenDesign address.
+func deriveBillingHost(host string) string {
+	if host == "amr-link.open-design.ai" {
+		return "amr-api.open-design.ai"
+	}
+	return host
+}
+
 // buildAutoChain rebuilds the AUTO chain snapshot from config.yaml rules
 // against the current providers. Slot model ids are normalized to the exact
 // configured spelling (case-insensitive match), otherwise ModelFor would fall
@@ -833,13 +835,35 @@ func (c *Core) Reload() error {
 // ---- adapter management (web UI) ----
 
 func (c *Core) ListAdapters() []lua.APIAdapter {
+	c.syncHiddenRefs()
 	return c.vm.ListAdapters()
 }
 
+// syncHiddenRefs 把"当前被源引用的适配器"同步给 VM。被引用的 hidden=true 适配器
+// 必须保持可见（源编辑器要回显当前值）；没被引用的隐藏适配器从列表消失。
+// 每次 ListAdapters 前调用：源集合是动态的（运行时可增删），推模型容易漏。
+func (c *Core) syncHiddenRefs() {
+	refs := []string{}
+	for _, s := range c.mergedSources() {
+		if s.Adapter != "" {
+			refs = append(refs, s.Adapter)
+		}
+	}
+	c.vm.SetHiddenRefs(refs)
+}
+
 func (c *Core) UploadAdapter(name, code string) error {
 	if name == "" {
 		return fmt.Errorf("adapter name required")
 	}
+	// Validate name SAFETY before touching the filesystem (R6). A path-traversal
+	// name ("../openai") must be rejected BEFORE WriteFile, or it would already
+	// have escaped the adapter dir by the time LoadAdapter re-checks. Built-in
+	// names are allowed here: an admin uploading openai.lua is the documented
+	// override, same as placing the file in adapter_dir by hand.
+	if err := lua.ValidateAdapterName(name, nil); err != nil {
+		return err
+	}
 	if err := os.MkdirAll(c.cfg.AdapterDir, 0755); err != nil {
 		return err
 	}
@@ -851,6 +875,13 @@ func (c *Core) UploadAdapter(name, code string) error {
 }
 
 func (c *Core) RemoveAdapter(name string) error {
+	// Validate the name so a traversal name cannot os.Remove a file outside the
+	// adapter dir. A reserved (built-in) name is a legitimate delete target in
+	// the running VM, so reserved-name rejection is NOT applied here — only the
+	// path-safety character set is.
+	if err := lua.ValidateAdapterName(name, nil); err != nil {
+		return err
+	}
 	path := filepath.Join(c.cfg.AdapterDir, name+".lua")
 	_ = os.Remove(path)
 	c.vm.RemoveAdapter(name)
@@ -974,63 +1005,6 @@ func (c *Core) Close() {
 	}
 }
 
-// applyBillingDSL compiles the configured billing profile and injects it into
-// the billing plugin.
-//
-// Configured pricing replaces the hand-written JSON blob that used to be PUT
-// through the state API. That path was used twice in production and both times
-// it "succeeded" while being wrong: once without the required `prices` wrapper
-// (silently wiping the accumulated totals), once with a peak-window shape the
-// plugin does not read (peak traffic billed at off-peak rates, no error). A
-// typed config with load-time validation turns both mistakes into startup
-// errors naming the file and field.
-//
-// A DSL with no profile, or one that fails to compile, leaves the plugin's
-// prices untouched: billing keeps running on whatever it had, and the problem
-// is logged rather than becoming a startup failure — the gateway must forward
-// even when its cost accounting is misconfigured.
-func (c *Core) applyBillingDSL() {
-	ps := c.plugins
-	if ps == nil || ps.Count() == 0 {
-		return
-	}
-	// The billing plugin may simply not be installed; a DSL without it is a
-	// no-op, not an error.
-	var have bool
-	for _, row := range ps.List() {
-		if row["name"] == "billing" && row["loaded"] == true {
-			have = true
-			break
-		}
-	}
-	if !have {
-		return
-	}
-	dsl := c.cfg.BillingDSL
-	if dsl == nil || len(dsl.Profiles) == 0 {
-		return
-	}
-	if err := dsl.Validate(); err != nil {
-		log.Printf("[billing] config invalid, profile not applied: %v", err)
-		return
-	}
-	profile := dsl.Resolve(dsl.Active)
-	if profile == nil {
-		log.Printf("[billing] no billing profile resolved (active=%q)", dsl.Active)
-		return
-	}
-	prices, err := billing.Compile(profile, c.cfg.Sources)
-	if err != nil {
-		log.Printf("[billing] profile %q not applied: %v", profile.ID, err)
-		return
-	}
-	if err := ps.SetState("billing", map[string]interface{}{"prices": prices}); err != nil {
-		log.Printf("[billing] applying profile %q: %v", profile.ID, err)
-		return
-	}
-	log.Printf("[billing] profile %q applied (%d rules, %d sources)", profile.ID, len(profile.Rules), len(c.cfg.Sources))
-}
-
 // ValidateScopeQuotas checks every scope entry's caps before they are stored.
 // A typo in a period must be rejected at write time rather than silently
 // becoming a never-resetting budget — the opposite of what was typed.

internal/core/core_test.go

@@ -5,7 +5,7 @@ import (
 	"sync"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // newTestConfig returns an in-memory config backed by temp files so
@@ -22,7 +22,7 @@ func newTestConfig(t *testing.T) *config.Config {
 		DefaultModel: "AUTO",
 		GatewayKeys:  []string{"sk-gw-test"},
 		Sources: []config.Source{
-			{Name: "s1", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai", Models: []config.Model{{ID: "gpt-4o", Priority: 10}}},
+			{Name: "s1", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai", Models: []config.Model{{ID: "gpt-4o"}}},
 		},
 	}
 }

internal/core/kernel_actions.go

@@ -0,0 +1,401 @@
+package core
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"modelrouter/internal/config"
+	"modelrouter/internal/lua"
+)
+
+// 内核动作实现（R4-c）
+//
+// 把 Core 的公开动作暴露给插件 Lua，走 lua.KernelAPI 接口注入（依赖方向是
+// core → lua，所以接口定义在 lua 包、实现放在这里）。
+//
+// # 为什么是"动作面"而不是"给插件一个 Core 指针"
+//
+// 直接暴露 *Core 会让插件能碰任何未导出的细节，而且每一次内核重构都会变成
+// 对插件作者的破坏性变更。动作面是一份**显式的契约**：每个动作有名字、危险级、
+// 超时预算，加动作是一次有意的行为，而不是顺手多导出一个方法。
+//
+// # 凭据边界
+//
+// 这里**不实现** list_keys / find_key。Core.ListKeys() 返回的是**明文**网关
+// 密钥，把它交给第三方 Lua 等于把部署的整个凭据面交出去。key 相关的读走掩码
+// 形态（create_key 的返回是唯一一次明文，且是"刚创建的那一个"）。
+var _ lua.KernelAPI = (*Core)(nil)
+
+// Actions implements lua.KernelAPI.
+//
+// The catalog lives in the lua package because permissions must be checkable even
+// when no Core is wired. This method exists so the interface is satisfied by the
+// real kernel rather than by a separate adapter type.
+func (c *Core) Actions() []lua.KernelAction { return lua.KernelActions() }
+
+// Invoke implements lua.KernelAPI. It dispatches one action.
+//
+// Every action is expected to honour ctx. The caller (lua.callKernel) gives up on
+// a deadline and returns an error to the plugin, but a kernel action that keeps
+// running afterwards would be an invisible leak, so the long ones (probe_all)
+// pass ctx down into the network layer.
+func (c *Core) Invoke(ctx context.Context, name string, args map[string]interface{}) (interface{}, error) {
+	switch name {
+	// ---- 只读 ----
+	case "status":
+		return c.actionStatus(), nil
+	case "list_sources":
+		return c.actionListSources(), nil
+	case "list_models":
+		return c.actionListModels(), nil
+	case "list_adapters":
+		return c.actionListAdapters(), nil
+	case "list_auto_rules":
+		return c.actionListAutoRules(), nil
+	case "adapter_owner":
+		return c.actionAdapterOwner(args)
+
+	// ---- 写 ----
+	case "reload":
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		return map[string]interface{}{"ok": true}, c.Reload()
+	case "reset_health":
+		// ResetHealth already no-ops on a nil registry (it ranges over
+		// Providers()), so this needs no guard beyond the nil check inside.
+		c.ResetHealth()
+		return map[string]interface{}{"ok": true}, nil
+	case "probe_all":
+		r := c.Registry()
+		if r == nil {
+			return nil, fmt.Errorf("the registry is not built yet")
+		}
+		r.ProbeAll(ctx)
+		if err := ctx.Err(); err != nil {
+			return nil, fmt.Errorf("probe_all was cut short: %w", err)
+		}
+		return map[string]interface{}{"ok": true}, nil
+	case "save_auto_rules":
+		return c.actionSaveAutoRules(args)
+	case "save_key_auto":
+		return c.actionSaveKeyAuto(args)
+	case "add_source":
+		return c.actionAddSource(args)
+	case "create_key":
+		return c.actionCreateKey(args)
+
+	// ---- 破坏性 ----
+	case "remove_source":
+		return c.actionRemoveSource(args)
+	case "delete_key":
+		return c.actionDeleteKey(args)
+	case "update_key":
+		return c.actionUpdateKey(args)
+	case "remove_adapter":
+		return c.actionRemoveAdapter(args)
+	case "set_plugin_enabled":
+		return c.actionSetPluginEnabled(args)
+	default:
+		return nil, fmt.Errorf("unknown kernel action %q", name)
+	}
+}
+
+// ---- 只读实现 ----
+
+func (c *Core) actionStatus() interface{} {
+	// A plugin can call this DURING LoadDir (its own __init or first hook), which
+	// runs before rebuildRegistry builds the registry. Returning an empty status
+	// rather than panicking keeps a load-time call from taking the process down;
+	// the plugin sees an empty list and the next call gives real data.
+	out := map[string]interface{}{
+		"sources":    []interface{}{},
+		"auto_slots": []interface{}{},
+		"default":    c.cfg.DefaultModel,
+		"ready":      false,
+	}
+	if r := c.Registry(); r != nil {
+		out["sources"] = r.Status()
+		out["ready"] = true
+	}
+	if slots := c.AutoSlotStates(); slots != nil {
+		out["auto_slots"] = slots
+	}
+	return out
+}
+
+// actionListSources returns the sources with api_key MASKED.
+//
+// Masking is not optional: a plugin has no legitimate need for another source's
+// key, and a plugin that wrote the source itself already has it in its own
+// credential vault.
+func (c *Core) actionListSources() interface{} {
+	srcs := c.Sources()
+	out := make([]map[string]interface{}, 0, len(srcs))
+	for _, s := range srcs {
+		models := make([]string, 0, len(s.Models))
+		for _, m := range s.Models {
+			models = append(models, m.ID)
+		}
+		out = append(out, map[string]interface{}{
+			"name":           s.Name,
+			"base_url":       s.BaseURL,
+			"adapter":        s.Adapter,
+			"models":         models,
+			"has_key":        s.APIKey != "" || s.APIKeyEnv != "",
+			"key_hint":       maskSecretHint(s.APIKey),
+			"max_concurrent": s.MaxConcurrent,
+			// rpm must be exposed too: it carries the per-model minimum gap, and
+			// max_concurrent alone cannot distinguish "serial WITH a gap" from
+			// "serial with no gap" — the latter still collides with the
+			// upstream's per-model concurrency window.
+			"rpm": s.RPM,
+		})
+	}
+	return out
+}
+
+// maskSecretHint shows only a short suffix so an operator (and a plugin) can tell
+// two keys apart without either getting the secret.
+func maskSecretHint(k string) string {
+	if k == "" {
+		return ""
+	}
+	if len(k) <= 4 {
+		return "***"
+	}
+	return "***" + k[len(k)-4:]
+}
+
+func (c *Core) actionListModels() interface{} {
+	if r := c.Registry(); r != nil {
+		return r.ModelList()
+	}
+	return []string{}
+}
+
+func (c *Core) actionListAdapters() interface{} {
+	return c.ListAdapters()
+}
+
+func (c *Core) actionListAutoRules() interface{} {
+	return map[string]interface{}{
+		"auto":       c.AutoRules(),
+		"auto_image": c.AutoImageRules(),
+	}
+}
+
+func (c *Core) actionAdapterOwner(args map[string]interface{}) (interface{}, error) {
+	name, _ := args["name"].(string)
+	if name == "" {
+		return nil, fmt.Errorf("adapter_owner requires a name")
+	}
+	owner := ""
+	if ps := c.Plugins(); ps != nil {
+		owner = ps.AdapterOwnerOf(name)
+	}
+	return map[string]interface{}{
+		"name":  name,
+		"owner": owner, // "" means operator-installed
+	}, nil
+}
+
+// ---- 写实现 ----
+
+func (c *Core) actionSaveAutoRules(args map[string]interface{}) (interface{}, error) {
+	entries, err := decodeScopes(args["entries"])
+	if err != nil {
+		return nil, err
+	}
+	if err := c.SaveAutoRules(entries); err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "count": len(entries)}, nil
+}
+
+func (c *Core) actionSaveKeyAuto(args map[string]interface{}) (interface{}, error) {
+	keyName, _ := args["key_name"].(string)
+	if keyName == "" {
+		return nil, fmt.Errorf("save_key_auto requires key_name")
+	}
+	entries, err := decodeScopes(args["entries"])
+	if err != nil {
+		return nil, err
+	}
+	if err := c.SaveKeyAuto(keyName, entries); err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "count": len(entries)}, nil
+}
+
+// actionAddSource is the capability that makes "plugin-driven upstream" real: a
+// plugin that has just obtained credentials can wire its endpoint in as a source
+// without the operator editing config.yaml.
+func (c *Core) actionAddSource(args map[string]interface{}) (interface{}, error) {
+	raw, ok := args["source"]
+	if !ok {
+		return nil, fmt.Errorf("add_source requires a source table")
+	}
+	b, err := json.Marshal(raw)
+	if err != nil {
+		return nil, fmt.Errorf("source is not encodable: %w", err)
+	}
+	var src config.Source
+	if err := json.Unmarshal(b, &src); err != nil {
+		return nil, fmt.Errorf("source shape is invalid: %w", err)
+	}
+	if strings.TrimSpace(src.Name) == "" {
+		return nil, fmt.Errorf("source requires a name")
+	}
+	if strings.TrimSpace(src.BaseURL) == "" {
+		return nil, fmt.Errorf("source %s requires a base_url", src.Name)
+	}
+	if err := c.AddSource(src); err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "name": src.Name}, nil
+}
+
+// actionCreateKey creates a gateway key.
+//
+// The plaintext key is returned ONCE, in this response — the same posture as the
+// admin API. It is not retrievable later, so a plugin that needs it must store it
+// in its own credential vault immediately.
+func (c *Core) actionCreateKey(args map[string]interface{}) (interface{}, error) {
+	name, _ := args["name"].(string)
+	role, _ := args["role"].(string)
+	if role == "" {
+		role = "user"
+	}
+	models, err := decodeScopes(args["models"])
+	if err != nil {
+		return nil, err
+	}
+	k, err := c.CreateKey(name, role, models, "created by plugin")
+	if err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{
+		"key":    k.Key,
+		"name":   k.Name,
+		"role":   k.Role,
+		"models": k.Models,
+	}, nil
+}
+
+// ---- 破坏性实现 ----
+
+func (c *Core) actionRemoveSource(args map[string]interface{}) (interface{}, error) {
+	name, _ := args["name"].(string)
+	if name == "" {
+		return nil, fmt.Errorf("remove_source requires a name")
+	}
+	// The last source is not removable: an empty registry turns every request
+	// into a 503, and the operator would have to fix it from the filesystem.
+	if len(c.Sources()) <= 1 {
+		return nil, fmt.Errorf("refusing to remove the last source: the gateway would have nothing to route to")
+	}
+	if err := c.RemoveSource(name); err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "name": name}, nil
+}
+
+func (c *Core) actionDeleteKey(args map[string]interface{}) (interface{}, error) {
+	key, _ := args["key"].(string)
+	if key == "" {
+		return nil, fmt.Errorf("delete_key requires the key to delete")
+	}
+	ok, err := c.DeleteKey(key)
+	if err != nil {
+		return nil, err
+	}
+	if !ok {
+		return nil, fmt.Errorf("no such key")
+	}
+	return map[string]interface{}{"ok": true}, nil
+}
+
+func (c *Core) actionUpdateKey(args map[string]interface{}) (interface{}, error) {
+	key, _ := args["key"].(string)
+	if key == "" {
+		return nil, fmt.Errorf("update_key requires the key")
+	}
+	name, _ := args["name"].(string)
+	role, _ := args["role"].(string)
+	models, err := decodeScopes(args["models"])
+	if err != nil {
+		return nil, err
+	}
+	k, err := c.UpdateKey(key, name, role, models, "")
+	if err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "name": k.Name, "role": k.Role}, nil
+}
+
+// actionRemoveAdapter deletes an adapter, but only one the CALLER plugin owns —
+// the ownership check lives in Plugins.RemoveAdapter (see the adapter bridge).
+// A plugin can always remove its own; removing an operator's or a built-in is
+// refused there, and that refusal is the security property.
+func (c *Core) actionRemoveAdapter(args map[string]interface{}) (interface{}, error) {
+	name, _ := args["name"].(string)
+	if name == "" {
+		return nil, fmt.Errorf("remove_adapter requires a name")
+	}
+	ps := c.Plugins()
+	if ps == nil {
+		return nil, fmt.Errorf("no plugins loaded")
+	}
+	plugin, _ := args["plugin"].(string)
+	if plugin == "" {
+		return nil, fmt.Errorf("remove_adapter needs the calling plugin's name")
+	}
+	if err := ps.RemoveAdapter(plugin, name); err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "name": name}, nil
+}
+
+func (c *Core) actionSetPluginEnabled(args map[string]interface{}) (interface{}, error) {
+	name, _ := args["name"].(string)
+	if name == "" {
+		return nil, fmt.Errorf("set_plugin_enabled requires a name")
+	}
+	enabled, ok := args["enabled"].(bool)
+	if !ok {
+		return nil, fmt.Errorf("set_plugin_enabled requires enabled (boolean)")
+	}
+	ps := c.Plugins()
+	if ps == nil {
+		return nil, fmt.Errorf("no plugins loaded")
+	}
+	if err := ps.SetEnabled(name, enabled); err != nil {
+		return nil, err
+	}
+	return map[string]interface{}{"ok": true, "name": name, "enabled": enabled}, nil
+}
+
+// decodeScopes converts a Lua-decoded array of model-scope tables into
+// []config.ModelScope. It accepts both an explicit array and a single table so a
+// one-entry chain does not need a wrapper.
+func decodeScopes(v interface{}) ([]config.ModelScope, error) {
+	if v == nil {
+		return nil, nil
+	}
+	b, err := json.Marshal(v)
+	if err != nil {
+		return nil, fmt.Errorf("entries are not encodable: %w", err)
+	}
+	var list []config.ModelScope
+	if err := json.Unmarshal(b, &list); err == nil {
+		return list, nil
+	}
+	var one config.ModelScope
+	if err := json.Unmarshal(b, &one); err != nil {
+		return nil, fmt.Errorf("entries shape is invalid: %w", err)
+	}
+	return []config.ModelScope{one}, nil
+}

internal/core/kernel_actions_test.go

@@ -0,0 +1,373 @@
+package core
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"modelrouter/internal/config"
+	"modelrouter/internal/lua"
+)
+
+// 本文件验证 R4-c/R4-d 的闭环：**插件自己驱动外部源的链接认证，然后把源接进内核**。
+//
+// 这是用户点明的"同类插件"能力（对照 dsh-our-free-model）：插件主导 OAuth 登录、
+// 存凭据、定时领额度、把拿到的端点注册成内核上游。此前 llmsproxy 的插件只能管
+// 自己的适配器，够不到内核。
+
+// kernelTestCore 起一个最小 core（一个源 + 一个 key）。
+func kernelTestCore(t *testing.T) *Core {
+	t.Helper()
+	dir := t.TempDir()
+	pluginDir := filepath.Join(dir, "plugins")
+	// 插件的自有存储位置由插件自己决定（读 config.yaml 的 plugin_dir，或
+	// MODELROUTER_PLUGIN_DIR 覆盖）。测试**必须**指向临时目录 —— 否则插件会读到
+	// /etc/modelrouter/config.yaml 并写进**生产插件目录**（真实事故：一次全量测试
+	// 把 26 个测试账号写进了生产 .channels-store.json）。
+	// 调用方已显式设定时尊重它：多次「重启」要共用同一目录才能测持久化。
+	if os.Getenv("MODELROUTER_PLUGIN_DIR") == "" {
+		t.Setenv("MODELROUTER_PLUGIN_DIR", pluginDir)
+	}
+	// ★ 显示器号与端口用**高段号**，避开生产。
+	//
+	// 真实事故（2026-10-15）：生产 unit 带 `PrivateTmp=yes`，它起的 Xvfb :91 的
+	// 锁文件在**服务自己的 /tmp** 里 —— 宿主上的测试看不到它，于是判"91 空闲"
+	// 并去抢，结果 `Xvfb 未就绪（/tmp/.X11-unix/X91 未出现）`，而生产上正跑着
+	// 一套浏览器栈。判据本身没错，错的是**两个 /tmp 互不可见**而号段重叠。
+	// 故测试整体挪到 300+。
+	if os.Getenv("MODELROUTER_XDISPLAY_BASE") == "" {
+		t.Setenv("MODELROUTER_XDISPLAY_BASE", "300")
+	}
+	cfg := &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		PluginDir:    pluginDir,
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		DefaultModel: "AUTO",
+		GatewayKeys:  []string{"sk-kernel-0001"},
+		Sources: []config.Source{{
+			Name: "primary", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai",
+			APIKey: "secret-primary-key", MaxConcurrent: 1,
+			Models: []config.Model{{ID: "m1", Kind: "chat"}},
+		}},
+	}
+	c, err := NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 测试结束**必须**把插件拉起的服务端浏览器栈收掉。
+	// 为何必需：callback 流（trae/gemini/有道）在 start_login 时就起一整套
+	// Xvfb+chromium+x11vnc+websockify，而绝大多数测试不会显式 stop_browser、
+	// 会话 TTL 也还没到 —— 于是一轮 core 全包跑完就攒出 7 套常驻进程
+	// （实测 websockify×7 + x11vnc×7 + 7 个 profile 目录），把显示器号与端口
+	// 段逐步耗光，后续测试报「chromium/websockify 端口未就绪」。
+	// 在这一处统一收掉，比让每个测试各自记得更可靠。
+	t.Cleanup(func() {
+		ps := c.Plugins()
+		if ps == nil {
+			return
+		}
+		for _, name := range []string{"channels"} {
+			_, _ = ps.InvokeHandler(name, "stop_browser", &lua.HandlerRequest{
+				Method: "POST", Role: "admin",
+			})
+		}
+	})
+	t.Cleanup(c.Close)
+	return c
+}
+
+// grant 给插件授权若干动作（模拟运维在设置页点选）。
+func grant(t *testing.T, c *Core, plugin string, actions ...string) {
+	t.Helper()
+	if err := c.Plugins().SetPluginKernelPermission(plugin, actions, false); err != nil {
+		t.Fatalf("grant %v to %s: %v", actions, plugin, err)
+	}
+}
+
+const kernelProbePlugin = `
+local plugin = { name = "kernelprobe", version = "1.0" }
+plugin.exports = { "read_status", "read_models", "read_sources", "try_delete_key", "probe_count", "call_reload" }
+plugin.kernel = { "status", "list_models", "reload" }
+
+function plugin.read_status()
+  local st, err = __llmsproxy_kernel.call("status", {})
+  if err then return { err = err } end
+  return { ready = st.ready, nsources = #(st.sources or {}) }
+end
+
+function plugin.read_models()
+  local m, err = __llmsproxy_kernel.call("list_models", {})
+  if err then return { err = err } end
+  return { models = m }
+end
+
+-- sources 动作返回的密钥必须是掩码的
+function plugin.read_sources()
+  local s, err = __llmsproxy_kernel.call("list_sources", {})
+  if err then return { err = err } end
+  return { sources = s }
+end
+
+-- 未授权的破坏性动作必须被拒
+function plugin.try_delete_key()
+  local _, err = __llmsproxy_kernel.call("delete_key", { key = "sk-kernel-0001" })
+  return { err = err }
+end
+
+-- 授权数量可见（插件据此优雅降级）
+function plugin.probe_count()
+  local acts = __llmsproxy_kernel.actions()
+  local granted = 0
+  for _, a in ipairs(acts) do
+    if a.granted then granted = granted + 1 end
+  end
+  return { total = #acts, granted = granted }
+end
+
+function plugin.call_reload()
+  local r, err = __llmsproxy_kernel.call("reload", {})
+  return { ok = r and r.ok, err = err }
+end
+return plugin
+`
+
+// TestPluginCallsKernelActions 验证插件能调内核动作，且权限边界成立。
+func TestPluginCallsKernelActions(t *testing.T) {
+	c := kernelTestCore(t)
+	if err := c.Plugins().LoadSource("kernelprobe", kernelProbePlugin); err != nil {
+		t.Fatal(err)
+	}
+	grant(t, c, "kernelprobe", "status", "list_models", "list_sources")
+
+	// 只读动作可用
+	out, err := c.Plugins().Call("kernelprobe", "read_status", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("read_status: %v", err)
+	}
+	m := out.(map[string]interface{})
+	if m["err"] != nil {
+		t.Fatalf("status 调用失败: %v", m["err"])
+	}
+	if m["ready"] != true {
+		t.Errorf("registry 应已就绪: %#v", m)
+	}
+	if n, _ := m["nsources"].(float64); n < 1 {
+		t.Errorf("应至少有 1 个源, 实得 %v", m["nsources"])
+	}
+
+	// 模型列表
+	out, err = c.Plugins().Call("kernelprobe", "read_models", map[string]interface{}{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if models, ok := out.(map[string]interface{})["models"].([]interface{}); !ok || len(models) == 0 {
+		t.Errorf("模型列表应为非空数组: %#v", out)
+	}
+}
+
+// TestKernelActionMasksSourceKeys 是安全断言：list_sources **不得**泄漏 api_key。
+func TestKernelActionMasksSourceKeys(t *testing.T) {
+	c := kernelTestCore(t)
+	if err := c.Plugins().LoadSource("kernelprobe", kernelProbePlugin); err != nil {
+		t.Fatal(err)
+	}
+	grant(t, c, "kernelprobe", "list_sources")
+
+	out, err := c.Plugins().Call("kernelprobe", "read_sources", map[string]interface{}{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 把整个返回值序列化后检查：明文密钥绝不能出现在里面
+	b, _ := json.Marshal(out)
+	if s := string(b); json.Valid(b) && contains(s, "secret-primary-key") {
+		t.Fatalf("list_sources 泄漏了源的明文 api_key: %s", s)
+	}
+	// 但应有可区分性（has_key + 掩码后缀）
+	if !contains(string(b), "***") {
+		t.Errorf("应给出掩码提示而不是什么都没有: %s", string(b))
+	}
+}
+
+// TestKernelActionDeniedWithoutGrant 是权限默认拒绝的断言。
+func TestKernelActionDeniedWithoutGrant(t *testing.T) {
+	c := kernelTestCore(t)
+	if err := c.Plugins().LoadSource("kernelprobe", kernelProbePlugin); err != nil {
+		t.Fatal(err)
+	}
+	// 故意不授权任何动作
+	out, err := c.Plugins().Call("kernelprobe", "try_delete_key", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("调用本身不该报错（拒绝应作为返回值）: %v", err)
+	}
+	m := out.(map[string]interface{})
+	if m["err"] == nil || m["err"] == "" {
+		t.Fatal("未授权的 delete_key 必须被拒")
+	}
+	t.Logf("拒绝信息: %v", m["err"])
+	// 拒绝信息要点名动作与插件（否则运维无法行动）
+	msg, _ := m["err"].(string)
+	if !contains(msg, "delete_key") || !contains(msg, "kernelprobe") {
+		t.Errorf("拒绝信息应点名动作与插件, 实得 %q", msg)
+	}
+	// 且密钥必须还在
+	if _, ok := c.FindKey("sk-kernel-0001"); !ok {
+		t.Fatal("被拒的动作不应真的执行")
+	}
+}
+
+// TestKernelActionsReportGranted 插件能看到自己拿到了哪些权限。
+func TestKernelActionsReportGranted(t *testing.T) {
+	c := kernelTestCore(t)
+	if err := c.Plugins().LoadSource("kernelprobe", kernelProbePlugin); err != nil {
+		t.Fatal(err)
+	}
+	grant(t, c, "kernelprobe", "status", "list_models")
+
+	out, err := c.Plugins().Call("kernelprobe", "probe_count", map[string]interface{}{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m := out.(map[string]interface{})
+	total, _ := m["total"].(float64)
+	granted, _ := m["granted"].(float64)
+	if total < 15 {
+		t.Errorf("动作总数应 >=15, 实得 %v", total)
+	}
+	if granted != 2 {
+		t.Errorf("授权数应为 2（status+list_models）, 实得 %v", granted)
+	}
+}
+
+// TestPluginRegistersSourceIntoKernel 是"同类插件"的核心闭环：
+// 插件通过内核动作把一个新的外部源接进网关。
+func TestPluginRegistersSourceIntoKernel(t *testing.T) {
+	c := kernelTestCore(t)
+	code := `
+local plugin = { name = "upstream", version = "1.0" }
+plugin.exports = { "add_my_source" }
+plugin.kernel = { "add_source", "list_sources" }
+
+function plugin.add_my_source(args)
+  local src = {
+    name = "plugin-provided",
+    base_url = "http://127.0.0.1:9/v1",
+    adapter = "openai",
+    api_key = "plugin-managed-token",
+    max_concurrent = 2,
+    models = { { id = "plugin-model", kind = "chat", priority = 10 } },
+  }
+  local r, err = __llmsproxy_kernel.call("add_source", { source = src })
+  if err then return { err = err } end
+  return { ok = r.ok, name = r.name }
+end
+return plugin
+`
+	if err := c.Plugins().LoadSource("upstream", code); err != nil {
+		t.Fatal(err)
+	}
+	grant(t, c, "upstream", "add_source", "list_sources")
+
+	out, err := c.Plugins().Call("upstream", "add_my_source", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("add_source: %v", err)
+	}
+	m := out.(map[string]interface{})
+	if m["err"] != nil {
+		t.Fatalf("插件注册源失败: %v", m["err"])
+	}
+	if m["name"] != "plugin-provided" {
+		t.Fatalf("返回的源名不对: %#v", m)
+	}
+
+	// 内核侧必须真的有了这个源 —— 这是"接进网关"的定义
+	found := false
+	for _, s := range c.Sources() {
+		if s.Name == "plugin-provided" {
+			found = true
+			if s.BaseURL != "http://127.0.0.1:9/v1" {
+				t.Errorf("base_url 未落库: %q", s.BaseURL)
+			}
+		}
+	}
+	if !found {
+		t.Fatal("插件注册的源没有出现在内核源列表里")
+	}
+	// 且模型已可路由
+	for _, id := range c.Registry().ModelList() {
+		if id == "plugin-model" {
+			return // 通过
+		}
+	}
+	t.Errorf("插件注册的模型未进入可路由列表: %v", c.Registry().ModelList())
+}
+
+// TestKernelPermissionPersists 授权必须落盘（否则重启即失效）。
+func TestKernelPermissionPersists(t *testing.T) {
+	c := kernelTestCore(t)
+	if err := c.Plugins().LoadSource("kernelprobe", kernelProbePlugin); err != nil {
+		t.Fatal(err)
+	}
+	grant(t, c, "kernelprobe", "status", "reload")
+
+	pdir := c.Config().PluginDir
+	path := filepath.Join(pdir, ".kernel-perms.json")
+	b, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("授权未落盘: %v", err)
+	}
+	if !contains(string(b), "reload") {
+		t.Errorf("落盘内容缺少动作名: %s", b)
+	}
+
+	// 模拟重启：新建 Plugins、加载授权、权限仍在
+	vm2 := lua.NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm2.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm2.Stop()
+	ps2 := lua.NewPlugins(vm2, pdir)
+	if err := ps2.LoadKernelPermissions(); err != nil {
+		t.Fatal(err)
+	}
+	perm := ps2.PluginKernelPermission("kernelprobe")
+	if len(perm.Actions) != 2 {
+		t.Fatalf("重启后授权应保留 2 项, 实得 %#v", perm.Actions)
+	}
+}
+
+// TestUnknownKernelActionRejected 拼错的动作名必须在授权时就被拒，
+// 而不是静默授予"什么都不允许"。
+func TestUnknownKernelActionRejected(t *testing.T) {
+	c := kernelTestCore(t)
+	if err := c.Plugins().LoadSource("kernelprobe", kernelProbePlugin); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Plugins().SetPluginKernelPermission("kernelprobe", []string{"no_such_action"}, false); err == nil {
+		t.Fatal("未知动作名应被拒")
+	}
+}
+
+// TestKernelActionSurfaceExcludesPlaintextKeys 是凭据边界的回归断言：
+// 动作面里绝不能出现返回明文密钥的动作（ListKeys/FindKey 被有意排除）。
+func TestKernelActionSurfaceExcludesPlaintextKeys(t *testing.T) {
+	for _, a := range lua.KernelActions() {
+		switch a.Name {
+		case "list_keys", "find_key", "get_key":
+			t.Fatalf("动作面出现了明文密钥读取动作 %q —— 这会把部署的凭据面交给插件", a.Name)
+		}
+	}
+}
+
+func contains(haystack, needle string) bool {
+	return len(needle) > 0 && len(haystack) >= len(needle) &&
+		func() bool {
+			for i := 0; i+len(needle) <= len(haystack); i++ {
+				if haystack[i:i+len(needle)] == needle {
+					return true
+				}
+			}
+			return false
+		}()
+}

internal/core/kernel_effect_test.go

@@ -0,0 +1,39 @@
+package core
+
+import "testing"
+
+// TestGrantedDestructiveActionTakesEffect 确认授权后的破坏性动作**真的执行**。
+//
+// 上一条测试只证明"不死锁、有返回"。一个把所有动作都变成空转的 dispatch 也能
+// 通过它 —— 所以这条单独断言副作用确实落地（密钥真的被删）。
+func TestGrantedDestructiveActionTakesEffect(t *testing.T) {
+	c := kernelTestCore(t)
+	code := `
+local plugin = { name = "eff", version = "1.0" }
+plugin.exports = { "del" }
+function plugin.del(args)
+  local r, err = __llmsproxy_kernel.call("delete_key", { key = args.key })
+  return { ok = r ~= nil, err = err }
+end
+return plugin
+`
+	if err := c.Plugins().LoadSource("eff", code); err != nil {
+		t.Fatal(err)
+	}
+	_ = c.Plugins().SetPluginKernelPermission("eff", []string{"delete_key"}, false)
+
+	if _, ok := c.FindKey("sk-kernel-0001"); !ok {
+		t.Fatal("前置条件：key 应存在")
+	}
+	out, err := c.Plugins().Call("eff", "del", map[string]interface{}{"key": "sk-kernel-0001"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m := out.(map[string]interface{})
+	if m["ok"] != true {
+		t.Fatalf("动作未成功: %#v", m)
+	}
+	if _, ok := c.FindKey("sk-kernel-0001"); ok {
+		t.Fatal("动作报告成功但副作用没发生 —— dispatch 是空转")
+	}
+}

internal/core/kernel_no_deadlock_test.go

@@ -0,0 +1,69 @@
+package core
+
+import (
+	"testing"
+	"time"
+
+	"modelrouter/internal/lua"
+)
+
+// TestNoKernelActionDeadlocksFromLua 对**每个**内核动作做同一条断言：
+// 在插件 Lua 调用（此时持有 p.mu）内执行时不得死锁。
+//
+// 为什么是"每个"而不是"那个出问题的"：最初只有 set_plugin_enabled 被发现自死锁
+// （它经 SetEnabled 重入 p.mu），但"是否还有别的动作也会重入"只能靠枚举回答。
+// 这条测试就是那次枚举的固化 —— 将来新增动作时如果也重入插件锁，它会失败。
+//
+// 同时它验证了"超时后副作用仍在幕后执行"这个更隐蔽的变体：那个缺陷的表现是
+// 插件收到 error、但动作其实生效了。断言方式只是"不死锁 + 返回"，因为死锁与
+// 迟到副作用是同一个根因（在错误的锁上下文里执行）。
+func TestNoKernelActionDeadlocksFromLua(t *testing.T) {
+	for _, a := range lua.KernelActions() {
+		a := a
+		t.Run(a.Name, func(t *testing.T) {
+			c := kernelTestCore(t)
+			code := `
+local plugin = { name = "probe_all_actions", version = "1.0" }
+plugin.exports = { "run" }
+function plugin.run(args)
+  local r, err = __llmsproxy_kernel.call(args.action, args.arg or {})
+  return { ok = r ~= nil, err = err }
+end
+return plugin
+`
+			if err := c.Plugins().LoadSource("probe_all_actions", code); err != nil {
+				t.Fatal(err)
+			}
+			// 授权所有动作（含破坏性），以便真正执行到
+			all := make([]string, 0)
+			for _, x := range lua.KernelActions() {
+				all = append(all, x.Name)
+			}
+			if err := c.Plugins().SetPluginKernelPermission("probe_all_actions", all, false); err != nil {
+				t.Fatal(err)
+			}
+			arg := map[string]interface{}{}
+			switch a.Name {
+			case "name", "adapter_owner":
+				arg["name"] = "primary"
+			}
+			if a.Name == "adapter_owner" {
+				arg = map[string]interface{}{"name": "openai"}
+			}
+			done := make(chan struct{})
+			go func() {
+				defer close(done)
+				_, err := c.Plugins().Call("probe_all_actions", "run",
+					map[string]interface{}{"action": a.Name, "arg": arg})
+				if err != nil {
+					t.Logf("%s 返回 err=%v", a.Name, err)
+				}
+			}()
+			select {
+			case <-done:
+			case <-time.After(8 * time.Second):
+				t.Fatalf("动作 %s 在 Lua 调用内死锁（deferred=%v danger=%s）", a.Name, a.Deferred, a.Danger)
+			}
+		})
+	}
+}

internal/core/key_auto_concurrency_test.go

@@ -7,7 +7,7 @@ import (
 	"testing"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // AutoChainFor sits on the request hot path and reads keyAutoChains with no
@@ -151,7 +151,7 @@ func TestAutoChainForConcurrentWithSourceChanges(t *testing.T) {
 			name := fmt.Sprintf("churn%d", n)
 			_ = c.AddSource(config.Source{
 				Name: name, BaseURL: "http://127.0.0.1:9/v1", Adapter: "openai",
-				Models: []config.Model{{ID: "m-" + name, Priority: 10}}},
+				Models: []config.Model{{ID: "m-" + name}}},
 			)
 			_ = c.RemoveSource(name)
 		}

internal/core/key_auto_test.go

@@ -4,8 +4,8 @@ import (
 	"path/filepath"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/scheduler"
+	"modelrouter/internal/scheduler"
 )
 
 // Per-key AUTO chains let an admin give one user their own scheduling chain
@@ -32,7 +32,7 @@ func perKeyTestConfig(t *testing.T) *config.Config {
 		GatewayKeys:  []string{"sk-gw-admin"},
 		Sources: []config.Source{
 			{Name: "s1", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai",
-				Models: []config.Model{{ID: "gpt-4o", Priority: 10}, {ID: "gpt-4o-mini", Priority: 5}}},
+				Models: []config.Model{{ID: "gpt-4o"}, {ID: "gpt-4o-mini"}}},
 		},
 		// Global chain points at gpt-4o.
 		Auto: []config.ModelScope{{Model: "gpt-4o", Source: "s1", Tier: 1}},
@@ -198,7 +198,7 @@ func TestKeyAutoChainsRebuildOnSourceChange(t *testing.T) {
 	// Add the source; the per-key chain must pick it up without a restart.
 	cfg.Sources = append(cfg.Sources, config.Source{
 		Name: "s2", BaseURL: "http://127.0.0.1:2/v1", Adapter: "openai",
-		Models: []config.Model{{ID: "new-model", Priority: 10}},
+		Models: []config.Model{{ID: "new-model"}},
 	})
 	if err := c.Reload(); err != nil {
 		t.Fatalf("Reload: %v", err)

internal/core/mutation_guard_test.go

@@ -0,0 +1,22 @@
+package core
+
+import (
+	"os"
+	"testing"
+	"time"
+)
+
+// 把测试进程的时区钉成 **UTC**。
+//
+// 为什么必需：Qoder 的每日活动按 **UTC+8** 结算，判定必须用算术平移而不是
+// 本机时区。但这台机器**本身就是 HKT(+0800)** —— 于是在真实环境里
+// 「os.date("!*t")（UTC 平移）」与「os.date("*t")（本机时区）」结果**完全相同**，
+// 判据无法区分二者（实测：把实现改成用本机时区，判据照样通过）。
+// 钉成 UTC 后两者相差 8 小时，误用本机时区的实现立刻暴露。
+//
+// 只影响测试进程，不改系统时区。
+func TestMain(m *testing.M) {
+	os.Setenv("TZ", "UTC")
+	time.Local = time.UTC
+	os.Exit(m.Run())
+}

internal/core/source_hosts_test.go

@@ -0,0 +1,108 @@
+package core
+
+import (
+	"net/url"
+	"testing"
+
+	"modelrouter/internal/config"
+)
+
+// The plugin outbound allowlist is built from the hosts of configured sources.
+// That rule is correct for every adapter whose credential and endpoint live on
+// the same host — which was all of them until OpenDesign.
+//
+// OpenDesign splits them: inference is amr-link.open-design.ai, billing is
+// amr-api.open-design.ai. The quota plugin derives the billing host from the
+// inference one, so the host it asks for is NOT the host the operator
+// configured. The request is then refused by the allowlist with a message that
+// reads like a plugin defect ("host ... is not an allowed outbound destination")
+// rather than a missing entry.
+//
+// This file pins the bridge: when a source looks like OpenDesign, the derived
+// billing host joins the allowlist.
+
+func TestSourceHostsIncludesConfiguredHosts(t *testing.T) {
+	cfg := &config.Config{Sources: []config.Source{
+		{Name: "a", BaseURL: "https://api.deepseek.com"},
+		{Name: "b", BaseURL: "https://amr-link.open-design.ai/v1"},
+		{Name: "empty", BaseURL: ""},
+		{Name: "bad", BaseURL: "://nonsense"},
+	}}
+	got := map[string]bool{}
+	for _, h := range sourceHosts(cfg) {
+		got[h] = true
+	}
+	for _, want := range []string{"api.deepseek.com", "amr-link.open-design.ai"} {
+		if !got[want] {
+			t.Errorf("sourceHosts missing configured host %q (allowlist would refuse "+
+				"every plugin request to it): %v", want, got)
+		}
+	}
+	// An empty/unparseable base_url must not abort the whole list.
+	if !got["api.deepseek.com"] {
+		t.Fatal("an unparseable sibling source dropped a valid host")
+	}
+}
+
+// TestSourceHostsAddsDerivedBillingHost covers the OpenDesign case at the level
+// that actually matters: the host the PLUGIN asks for must be in the allowlist.
+func TestSourceHostsAddsDerivedBillingHost(t *testing.T) {
+	cfg := &config.Config{Sources: []config.Source{
+		{Name: "vela", BaseURL: "https://amr-link.open-design.ai/v1"},
+	}}
+	got := map[string]bool{}
+	for _, h := range sourceHosts(cfg) {
+		got[h] = true
+	}
+	if !got["amr-link.open-design.ai"] {
+		t.Fatalf("inference host missing: %v", got)
+	}
+	if !got["amr-api.open-design.ai"] {
+		t.Fatalf("the billing host the quota plugin derives is not in the allowlist, "+
+			"so every OpenDesign balance probe is refused before it leaves the "+
+			"process: %v", got)
+	}
+}
+
+// A deployment with no OpenDesign source must not silently gain its billing
+// host: an allowlist entry is a grant, and widening it for everyone would let an
+// unrelated plugin reach a host the operator never configured.
+func TestSourceHostsDoesNotGrantUnconfiguredBillingHost(t *testing.T) {
+	cfg := &config.Config{Sources: []config.Source{
+		{Name: "ds", BaseURL: "https://api.deepseek.com"},
+	}}
+	for _, h := range sourceHosts(cfg) {
+		if h == "amr-api.open-design.ai" {
+			t.Fatal("the OpenDesign billing host was granted to a deployment with " +
+				"no OpenDesign source")
+		}
+	}
+}
+
+// The derivation has to agree with the Lua side. Keep the two spellings of the
+// rule in one place so a change to one is visible when reading the other.
+func TestSourceHostsDerivationMatchesPluginRule(t *testing.T) {
+	// what the Lua adapter does: origin:gsub("^amr%-link%.", "amr-api.")
+	const in = "amr-link.open-design.ai"
+	want := "amr-api.open-design.ai"
+	got := deriveBillingHost(in)
+	if got != want {
+		t.Fatalf("deriveBillingHost(%q) = %q, want %q — the Go allowlist and the "+
+			"Lua probe must agree on the host or the request is refused", in, got, want)
+	}
+	// A host that is not amr-link stays untouched, so this cannot accidentally
+	// rewrite an unrelated provider into an OpenDesign address.
+	for _, other := range []string{"api.deepseek.com", "amr-link.example.com", "xamr-link.a"} {
+		if deriveBillingHost(other) != other {
+			t.Errorf("deriveBillingHost(%q) = %q, want it unchanged", other, deriveBillingHost(other))
+		}
+	}
+	// url.Parse round-trip keeps the helper honest about what it receives.
+	u, err := url.Parse("https://amr-link.open-design.ai/v1")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if deriveBillingHost(u.Hostname()) != "amr-api.open-design.ai" {
+		t.Fatal("derivation failed on a real base_url hostname")
+	}
+}

internal/core/startup_secrets_test.go

@@ -6,7 +6,7 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // Startup must be idempotent, and every credential consumer must see PLAINTEXT.
@@ -42,7 +42,7 @@ func writeSealedInstall(t *testing.T, dir string) string {
 		GatewayKeys:  []string{"sk-gw-seed-me"},
 		Sources: []config.Source{{
 			Name: "up", BaseURL: "http://127.0.0.1:1/v1", APIKey: "sk-upstream-secret",
-			Adapter: "openai", Models: []config.Model{{ID: "gpt-4o", Priority: 10}},
+			Adapter: "openai", Models: []config.Model{{ID: "gpt-4o"}},
 		}},
 	}
 	c, err := NewFromConfig(cfg)
@@ -76,7 +76,7 @@ func writeSealedInstallNoSeed(t *testing.T, dir string) string {
 		DefaultModel: "AUTO",
 		Sources: []config.Source{{
 			Name: "up", BaseURL: "http://127.0.0.1:1/v1", APIKey: "sk-upstream-secret",
-			Adapter: "openai", Models: []config.Model{{ID: "gpt-4o", Priority: 10}},
+			Adapter: "openai", Models: []config.Model{{ID: "gpt-4o"}},
 		}},
 	}
 	c, err := NewFromConfig(cfg)

internal/core/trae_derive_expected.json

@@ -0,0 +1,44 @@
+[
+ {
+  "u": "12345",
+  "devid": "707304999019651",
+  "market": "5daa819c-8425-4bae-88b7-2bbfec3c25ba",
+  "sess": "e5b8dfc5830ce0bdb141f7177effb90a7cf8227bb23b930a427a8550bf7357df"
+ },
+ {
+  "u": "UID-abc",
+  "devid": "579362202930608",
+  "market": "a5388fc1-5969-4666-bc74-be9a11f69567",
+  "sess": "a47da2e13fe16d49dc859fbed58d1d99e0faef5d116acafc64a2d26013d17a1f"
+ },
+ {
+  "u": "900000001",
+  "devid": "764948639142479",
+  "market": "e449daa5-b0c0-4f35-918e-8bb01b33f179",
+  "sess": "fdd37c37d1cc950439c25467e6dca75fcc84a2b0d294aa910387a9a965645650"
+ },
+ {
+  "u": "x",
+  "devid": "268405722745284",
+  "market": "4f2427c5-0583-4f82-8e18-4e38566dd857",
+  "sess": "79be40fff0fbaee572d1ba3ab13a47af9d2f4d943115ef512c56e42132e23a66"
+ },
+ {
+  "u": "",
+  "devid": "170789623621479",
+  "market": "b32897c1-f9c0-4ad2-b2b4-725455fffb6b",
+  "sess": "4a61a6b4c4bcbf1040bf0edf3df2b3e7d76a956f3ef81ef8f148eadcc03d25cc"
+ },
+ {
+  "u": "用户1",
+  "devid": "263026789643405",
+  "market": "1c1b4e1f-9d7a-4698-9f34-b057d9a74b0b",
+  "sess": "a13dbde68977eee49818e2e2fb7d3bbb1ffb4c807a6e6f1831168a84adea0dfe"
+ },
+ {
+  "u": "1234567890123456789012345",
+  "devid": "358669731114854",
+  "market": "449882ec-9a69-4a7a-b9ab-25aa59b28c50",
+  "sess": "305eb96fe14a194386c0fd4148d4919870ed803b69cf75296b53ea1f41e4c574"
+ }
+]

internal/gateway/api.go

@@ -11,7 +11,7 @@ import (
 	"strings"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 type adapterPayload struct {
@@ -215,7 +215,7 @@ const keepExistingAPIKey = "__KEEP__"
 //   - the explicit mask keepExistingAPIKey — always inherit.
 //   - the empty string — inherit only for a source that already has a
 //     credential. Empty is a legitimate value for a *new* credential-less
-//     source (dshcodebuddy uses "dsh-local-noauth", local servers use ""),
+//     source (some local runtimes use a no-auth sentinel, local servers ""),
 //     so a fresh source with an empty key must stay empty rather than inherit
 //     nothing and fail differently.
 func (g *Gateway) resolveAPIKey(name, want string, exists bool) string {
@@ -470,8 +470,8 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 			// the all-time view showed the truth, which read as "the gateway is
 			// idle" precisely when it was busiest. A calendar window cannot
 			// filter it, so it travels with every response unfiltered.
-			"active": g.stats.Active(),
+			"active":  g.stats.Active(),
-			"records":   page.Records,
+			"records": page.Records,
 			// The paging cursor has to travel with the first page, otherwise
 			// the scroll handler has nothing to page from and the table is
 			// stuck at one screen forever.
@@ -488,7 +488,7 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 		if to == 0 {
 			to = time.Now().UnixMilli()
 		}
-		csvHeaders(w, "llmsproxy-requests.csv")
+		csvHeaders(w, "modelrouter-requests.csv")
 		cw := csv.NewWriter(w)
 		names := map[string]string{}
 		for _, k := range g.core.ListKeys() {
@@ -537,7 +537,7 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	if r.URL.Query().Get("export") == "keys-csv" {
-		csvHeaders(w, "llmsproxy-keys.csv")
+		csvHeaders(w, "modelrouter-keys.csv")
 		cw := csv.NewWriter(w)
 		_ = cw.Write([]string{"key", "key_name", "role", "models", "total_requests", "success_requests", "failed_requests", "prompt_tokens", "completion_tokens", "total_tokens", "avg_latency_ms", "max_latency_ms", "created_at"})
 		keyFilter := exportKey(r)

internal/gateway/apiv1.go

@@ -31,7 +31,7 @@ import (
 	"sort"
 	"strings"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // apiV1Routes dispatches /api/v1/*. Kept separate from routes() so the original
@@ -144,7 +144,7 @@ func (g *Gateway) apiV1Index(w http.ResponseWriter, r *http.Request) {
 	}
 	index := map[string]interface{}{
 		"api_version": "v1",
-		"description": "llmsproxy management API. Authenticate with a gateway key: " +
+		"description": "modelrouter management API. Authenticate with a gateway key: " +
 			"'Authorization: Bearer <key>', '?api_key=<key>', or the gw_key cookie. " +
 			"Read endpoints accept any role; mutations require an admin key.",
 		"discovery": []ep{
@@ -189,6 +189,20 @@ func (g *Gateway) apiV1Index(w http.ResponseWriter, r *http.Request) {
 			{Method: "GET", Path: "/api/stats", Auth: "any", Summary: "usage aggregates"},
 			{Method: "GET", Path: "/api/stats/records", Auth: "any", Summary: "paged request records"},
 			{Method: "GET", Path: "/api/adapters", Auth: "admin", Summary: "installed Lua adapters"},
+			{Method: "POST", Path: "/api/sources/discover", Auth: "admin",
+				Summary: "probe a base_url+api_key for its model catalog",
+				// Probing is read-only with respect to configuration: it writes
+				// nothing, which is why it is safe to call repeatedly while the
+				// operator is still typing a key.
+				WriteEffect: "none",
+			},
+			{Method: "POST", Path: "/api/sources/discover/apply", Auth: "admin",
+				Summary: "probe and write the source in one step",
+				// Requires replace_models=true when the source already exists,
+				// so re-running discovery cannot silently discard a hand-tuned
+				// model list. An empty api_key keeps the stored credential.
+				WriteEffect: "creates or replaces a source and its model list",
+			},
 			{Method: "GET", Path: "/api/source_templates", Auth: "admin", Summary: "source templates"},
 			{Method: "POST", Path: "/v1/chat/completions", Auth: "any", Summary: "OpenAI-compatible inference"},
 		},

internal/gateway/apiv1_test.go

@@ -7,7 +7,7 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // v1Gateway builds a gateway with one source and two keys, and returns the

internal/gateway/billing_rules_api.go

@@ -1,282 +0,0 @@
-package gateway
-
-import (
-	"encoding/json"
-	"net/http"
-
-	"llmsproxy/internal/billing"
-	"llmsproxy/internal/config"
-)
-
-// Billing rules API.
-//
-//	GET  /api/plugins/billing/rules    the active profile's rules + all profiles
-//	PUT  /api/plugins/billing/rules    replace the whole DSL and re-inject prices
-//	POST /api/plugins/billing/rules    { profile, rule } append one rule
-//	DELETE /api/plugins/billing/rules  { profile, url } remove one rule
-//
-// Why a separate endpoint instead of the generic plugin state: prices are
-// plugin state, but they are also a durable, reviewable CONFIGURATION. Routing
-// them through /state would let an admin key PUT arbitrary plugin state and
-// silently reset the accumulated accounting. Here the server owns the shape:
-// it validates the DSL, writes it back to config.yaml, recompiles, and hands
-// the plugin its prices table. The operator edits rules; the plugin keeps
-// numbers. The two are never mixed in one payload.
-//
-// Editing means "what the operator typed is what config.yaml holds". A
-// billing mistake found next week must be traceable to a reviewable file, not
-// to a blob in the plugin's state sidecar.
-func (g *Gateway) handleBillingRules(w http.ResponseWriter, r *http.Request) {
-	if reqRole(r.Context()) != "admin" {
-		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
-		return
-	}
-	switch r.Method {
-	case http.MethodGet:
-		g.getBillingRules(w)
-	case http.MethodPut:
-		g.putBillingRules(w, r)
-	case http.MethodPost:
-		g.addBillingRule(w, r)
-	case http.MethodDelete:
-		g.delBillingRule(w, r)
-	default:
-		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET/PUT/POST/DELETE")
-	}
-}
-
-// billingRulesPayload is the UI-facing view: the profiles as declared, plus
-// which one is active and the source URLs they can match — the editor needs
-// the URL list to offer suggestions, and an operator typing a URL that matches
-// no source is exactly the silent-no-pricing failure this feature exists to
-// remove.
-type billingRulesPayload struct {
-	DSL      *config.BillingDSL `json:"billing"`
-	Active   string             `json:"active"`
-	URLs     []string           `json:"urls"`
-	Warnings []string           `json:"warnings"`
-}
-
-func (g *Gateway) billingRulesPayload() billingRulesPayload {
-	out := billingRulesPayload{}
-	cfg := g.core.Config()
-	if cfg != nil && cfg.BillingDSL != nil {
-		// Copy so the response cannot be mutated back through the pointer.
-		cp := *cfg.BillingDSL
-		out.DSL = &cp
-		if p := cp.Resolve(""); p != nil {
-			out.Active = p.ID
-		}
-	}
-	for _, s := range g.sourceURLs() {
-		out.URLs = append(out.URLs, s)
-	}
-	out.Warnings = g.billingRuleWarnings()
-	return out
-}
-
-func (g *Gateway) sourceURLs() []string {
-	cfg := g.core.Config()
-	if cfg == nil {
-		return nil
-	}
-	seen := map[string]bool{}
-	var out []string
-	for _, s := range cfg.Sources {
-		if s.BaseURL == "" || seen[s.BaseURL] {
-			continue
-		}
-		seen[s.BaseURL] = true
-		out = append(out, s.BaseURL)
-	}
-	return out
-}
-
-// billingRuleWarnings reports rules that match no configured source. This is
-// the silent killer of hand-written price tables: the rule parses, validates,
-// and prices nothing at all, so every request on that URL lands in unpriced.
-// Saying so out loud is the difference between a five-second fix and an
-// afternoon wondering why the bill is zero.
-func (g *Gateway) billingRuleWarnings() []string {
-	cfg := g.core.Config()
-	if cfg == nil || cfg.BillingDSL == nil {
-		return nil
-	}
-	urls := map[string]bool{}
-	for _, u := range g.sourceURLs() {
-		urls[u] = true
-	}
-	var warns []string
-	for _, p := range cfg.BillingDSL.Profiles {
-		for _, r := range p.Rules {
-			if r.URL == "*" || urls[r.URL] {
-				continue
-			}
-			warns = append(warns, "profile "+p.ID+": rule url "+r.URL+" matches no configured source")
-		}
-	}
-	return warns
-}
-
-func (g *Gateway) getBillingRules(w http.ResponseWriter) {
-	writeJSON(w, http.StatusOK, g.billingRulesPayload())
-}
-
-func (g *Gateway) putBillingRules(w http.ResponseWriter, r *http.Request) {
-	var body struct {
-		Billing *config.BillingDSL `json:"billing"`
-	}
-	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
-		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
-		return
-	}
-	if body.Billing == nil {
-		writeError(w, http.StatusBadRequest, "invalid_request", "billing is required")
-		return
-	}
-	g.applyBillingDSLUpdate(w, body.Billing)
-}
-
-func (g *Gateway) addBillingRule(w http.ResponseWriter, r *http.Request) {
-	var body struct {
-		Profile string             `json:"profile"`
-		Rule    config.BillingRule `json:"rule"`
-	}
-	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
-		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
-		return
-	}
-	cfg := g.core.Config()
-	if cfg == nil || cfg.BillingDSL == nil || len(cfg.BillingDSL.Profiles) == 0 {
-		writeError(w, http.StatusBadRequest, "no_profile", "no billing profiles configured; create one first")
-		return
-	}
-	next := cloneBillingDSL(cfg.BillingDSL)
-	idx := profileIndex(next, body.Profile)
-	if idx < 0 {
-		writeError(w, http.StatusBadRequest, "unknown_profile", "no such profile: "+body.Profile)
-		return
-	}
-	p := &next.Profiles[idx]
-	// Replace rather than append when the URL is already declared: two rules
-	// for one URL would silently make the first unreachable (first match wins),
-	// which is the exact ambiguity an editor should not be able to create.
-	replaced := false
-	for i := range p.Rules {
-		if p.Rules[i].URL == body.Rule.URL {
-			p.Rules[i] = body.Rule
-			replaced = true
-			break
-		}
-	}
-	if !replaced {
-		p.Rules = append(p.Rules, body.Rule)
-	}
-	g.applyBillingDSLUpdate(w, next)
-}
-
-func (g *Gateway) delBillingRule(w http.ResponseWriter, r *http.Request) {
-	var body struct {
-		Profile string `json:"profile"`
-		URL     string `json:"url"`
-	}
-	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
-		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
-		return
-	}
-	cfg := g.core.Config()
-	if cfg == nil || cfg.BillingDSL == nil {
-		writeError(w, http.StatusBadRequest, "no_profile", "no billing profiles configured")
-		return
-	}
-	next := cloneBillingDSL(cfg.BillingDSL)
-	idx := profileIndex(next, body.Profile)
-	if idx < 0 {
-		writeError(w, http.StatusBadRequest, "unknown_profile", "no such profile: "+body.Profile)
-		return
-	}
-	p := &next.Profiles[idx]
-	out := p.Rules[:0]
-	found := false
-	for _, rule := range p.Rules {
-		if rule.URL == body.URL {
-			found = true
-			continue
-		}
-		out = append(out, rule)
-	}
-	if !found {
-		writeError(w, http.StatusNotFound, "not_found", "no rule for url "+body.URL)
-		return
-	}
-	p.Rules = out
-	g.applyBillingDSLUpdate(w, next)
-}
-
-// applyBillingDSLUpdate is the single write path: validate, persist to
-// config.yaml, recompile, inject, and only then report success. If the config
-// cannot be written the change is NOT applied in memory either — a rules editor
-// that says "saved" and loses the edit on the next restart is worse than one
-// that refuses.
-func (g *Gateway) applyBillingDSLUpdate(w http.ResponseWriter, next *config.BillingDSL) {
-	if err := next.Validate(); err != nil {
-		writeError(w, http.StatusBadRequest, "invalid_rules", err.Error())
-		return
-	}
-	cfg := g.core.Config()
-	if cfg == nil {
-		writeError(w, http.StatusInternalServerError, "no_config", "no config loaded")
-		return
-	}
-	// Assign BEFORE saving. The previous order saved first and assigned after,
-	// so every rule edit reported success while writing a config.yaml with no
-	// billing section at all — the operator's edit vanished on restart and
-	// nothing in the API response said so.
-	prev := cfg.BillingDSL
-	cfg.BillingDSL = next
-
-	// Compile before persisting: a profile that cannot be turned into prices
-	// must not reach the config file, or the next restart fails to load the
-	// very rules the editor just accepted.
-	prices, err := billing.CompileOpts(next.Resolve(next.Active), cfg.Sources, true)
-	if err != nil {
-		cfg.BillingDSL = prev // no half-applied state
-		writeError(w, http.StatusBadRequest, "compile_failed", err.Error())
-		return
-	}
-	if err := cfg.Save(); err != nil {
-		cfg.BillingDSL = prev
-		writeError(w, http.StatusInternalServerError, "persist_failed", err.Error())
-		return
-	}
-	ps := g.core.Plugins()
-	if ps != nil {
-		if err := ps.SetState("billing", map[string]interface{}{"prices": prices}); err != nil {
-			writeError(w, http.StatusBadRequest, "plugin_error", err.Error())
-			return
-		}
-	}
-	p := g.billingRulesPayload()
-	writeJSON(w, http.StatusOK, map[string]interface{}{
-		"ok": true, "billing": p.DSL, "active": p.Active, "warnings": p.Warnings,
-	})
-}
-
-func cloneBillingDSL(d *config.BillingDSL) *config.BillingDSL {
-	out := &config.BillingDSL{Active: d.Active}
-	out.Profiles = make([]config.BillingProfile, len(d.Profiles))
-	copy(out.Profiles, d.Profiles)
-	return out
-}
-
-func profileIndex(d *config.BillingDSL, id string) int {
-	if id == "" {
-		return -1
-	}
-	for i, p := range d.Profiles {
-		if p.ID == id {
-			return i
-		}
-	}
-	return -1
-}

internal/gateway/billing_rules_api_test.go

@@ -1,281 +0,0 @@
-package gateway
-
-import (
-	"encoding/json"
-	"net/http"
-	"net/http/httptest"
-	"os"
-	"sort"
-	"strings"
-	"testing"
-
-	"llmsproxy/internal/config"
-)
-
-// rulesGateway needs a real billing plugin (the API injects prices into it)
-// AND two sources whose base_urls the rules will match, so it cannot reuse
-// either existing helper: gatewayWithBilling has no sources, and
-// newTestGateway loads no plugins.
-func rulesGateway(t *testing.T) *Gateway {
-	t.Helper()
-	g := gatewayWithBilling(t)
-	cfg := g.core.Config()
-	cfg.Sources = []config.Source{
-		{Name: "localzen", BaseURL: "https://free.example.com/v1"},
-		{Name: "commandcode", BaseURL: "https://api.commandcode.ai/v1"},
-	}
-	return g
-}
-
-func rulesDo(t *testing.T, g *Gateway, method, body string) *httptest.ResponseRecorder {
-	t.Helper()
-	return doReq(t, g, method, "/api/plugins/billing/rules", body)
-}
-
-// TestBillingRulesCRUD covers the whole editor loop against the real HTTP
-// surface: read, add, replace-by-URL, delete. The replace-by-URL case is the
-// one worth pinning — two rules for one URL would make the first unreachable
-// (first match wins) without any error, which is exactly the kind of silent
-// ambiguity an editor must not be able to create.
-func TestBillingRulesCRUD(t *testing.T) {
-	g := rulesGateway(t)
-	seed := `{"billing":{"active":"p1","profiles":[{"id":"p1","rules":[
-      {"url":"https://free.example.com/v1","mode":"free"}]}]}}`
-	if rr := rulesDo(t, g, http.MethodPut, seed); rr.Code != http.StatusOK {
-		t.Fatalf("seed rules = %d: %s", rr.Code, rr.Body.String())
-	}
-
-	// Add a token-priced rule for the second URL.
-	add := `{"profile":"p1","rule":{"url":"https://api.commandcode.ai/v1","mode":"token",
-      "models":{"deepseek-v4.1-flash":{"prompt":"0.15","completion":"0.60"}}}}`
-	if rr := rulesDo(t, g, http.MethodPost, add); rr.Code != http.StatusOK {
-		t.Fatalf("add rule = %d: %s", rr.Code, rr.Body.String())
-	}
-	var got billingRulesPayload
-	if err := json.Unmarshal(rulesDo(t, g, http.MethodGet, "").Body.Bytes(), &got); err != nil {
-		t.Fatal(err)
-	}
-	if len(got.DSL.Profiles[0].Rules) != 2 {
-		t.Fatalf("after add there are %d rules, want 2", len(got.DSL.Profiles[0].Rules))
-	}
-
-	// Adding the SAME url again must replace, not append.
-	if rr := rulesDo(t, g, http.MethodPost, `{"profile":"p1","rule":{"url":"https://api.commandcode.ai/v1","mode":"free"}}`); rr.Code != http.StatusOK {
-		t.Fatalf("re-add same url = %d: %s", rr.Code, rr.Body.String())
-	}
-	if err := json.Unmarshal(rulesDo(t, g, http.MethodGet, "").Body.Bytes(), &got); err != nil {
-		t.Fatal(err)
-	}
-	if n := len(got.DSL.Profiles[0].Rules); n != 2 {
-		t.Fatalf("★ re-adding an existing url appended instead of replacing: %d rules", n)
-	}
-	var mode string
-	for _, r := range got.DSL.Profiles[0].Rules {
-		if r.URL == "https://api.commandcode.ai/v1" {
-			mode = r.Mode
-		}
-	}
-	if mode != "free" {
-		t.Errorf("re-added rule mode = %q, want free (the replacement must win)", mode)
-	}
-
-	// Delete by URL.
-	if rr := rulesDo(t, g, http.MethodDelete, `{"profile":"p1","url":"https://api.commandcode.ai/v1"}`); rr.Code != http.StatusOK {
-		t.Fatalf("delete rule = %d: %s", rr.Code, rr.Body.String())
-	}
-	if err := json.Unmarshal(rulesDo(t, g, http.MethodGet, "").Body.Bytes(), &got); err != nil {
-		t.Fatal(err)
-	}
-	if len(got.DSL.Profiles[0].Rules) != 1 {
-		t.Errorf("after delete there are %d rules, want 1", len(got.DSL.Profiles[0].Rules))
-	}
-}
-
-// TestBillingRulesRejectUnknownURLWithWarning is the silent-failure guard: a
-// rule whose URL matches no source prices nothing, so every request on it
-// falls into unpriced. The API must say so instead of accepting it silently.
-func TestBillingRulesRejectUnknownURLWithWarning(t *testing.T) {
-	g := rulesGateway(t)
-	if rr := rulesDo(t, g, http.MethodPut, `{"billing":{"active":"p","profiles":[{"id":"p",
-      "rules":[{"url":"https://typo.example.com/v1","mode":"free"}]}]}}`); rr.Code != http.StatusOK {
-		t.Fatalf("seed = %d: %s", rr.Code, rr.Body.String())
-	}
-	var got billingRulesPayload
-	if err := json.Unmarshal(rulesDo(t, g, http.MethodGet, "").Body.Bytes(), &got); err != nil {
-		t.Fatal(err)
-	}
-	if len(got.Warnings) == 0 {
-		t.Fatal("★ a rule matching no configured source was accepted with no warning — " +
-			"it will price nothing and every request will be recorded as unpriced")
-	}
-	if !strings.Contains(got.Warnings[0], "typo.example.com") {
-		t.Errorf("warning does not name the offending url: %v", got.Warnings)
-	}
-	// The editor needs the real URL list to offer suggestions.
-	if len(got.URLs) != 2 {
-		t.Errorf("urls offered to the editor = %v, want the 2 configured base_urls", got.URLs)
-	}
-}
-
-// TestBillingRulesInvalidIsRejectedNotPersisted checks the write path refuses
-// bad input and leaves the previous rules in place — an editor that clears the
-// table on a typo loses the price table.
-func TestBillingRulesInvalidIsRejectedNotPersisted(t *testing.T) {
-	g := rulesGateway(t)
-	seed := `{"billing":{"active":"p","profiles":[{"id":"p","rules":[
-      {"url":"https://free.example.com/v1","mode":"free"}]}]}}`
-	if rr := rulesDo(t, g, http.MethodPut, seed); rr.Code != http.StatusOK {
-		t.Fatalf("seed = %d: %s", rr.Code, rr.Body.String())
-	}
-	// An unknown mode must be refused.
-	if rr := rulesDo(t, g, http.MethodPost, `{"profile":"p","rule":{"url":"https://x/v1","mode":"banana"}}`); rr.Code != http.StatusBadRequest {
-		t.Errorf("invalid mode accepted with %d, want 400", rr.Code)
-	}
-	var got billingRulesPayload
-	if err := json.Unmarshal(rulesDo(t, g, http.MethodGet, "").Body.Bytes(), &got); err != nil {
-		t.Fatal(err)
-	}
-	if len(got.DSL.Profiles[0].Rules) != 1 {
-		t.Errorf("★ a rejected rule was persisted anyway: %d rules", len(got.DSL.Profiles[0].Rules))
-	}
-	// An unparseable price must be refused too, not silently priced at zero.
-	if rr := rulesDo(t, g, http.MethodPost, `{"profile":"p","rule":{"url":"https://y/v1","mode":"token",
-      "models":{"m":{"prompt":"free","completion":"0.6"}}}}`); rr.Code != http.StatusBadRequest {
-		t.Errorf("non-numeric price accepted with %d, want 400", rr.Code)
-	}
-}
-
-// TestBillingRulesPersistAndReachThePlugin is the point of the whole
-// endpoint, and mutation verification showed the earlier CRUD tests missed it
-// twice: dropping cfg.Save() and dropping the SetState injection both left
-// every test green. Both failures are invisible in a GET — the response echoes
-// the in-memory rules either way — so this test checks the two consequences
-// that actually matter:
-//
-//  1. persistence: the rule must be in the config file on disk, because an
-//     edit that evaporates on restart is a lie the UI told the operator;
-//  2. effect: the billing plugin must now PRICE the source (its published
-//     prices table has a non-zero entry), because a rule that is stored but
-//     never injected prices nothing at all.
-func TestBillingRulesPersistAndReachThePlugin(t *testing.T) {
-	g := rulesGateway(t)
-	cfg := g.core.Config()
-
-	seed := `{"billing":{"active":"p","profiles":[{"id":"p","rules":[
-      {"url":"https://free.example.com/v1","mode":"free"},
-      {"url":"https://api.commandcode.ai/v1","mode":"token",
-       "models":{"deepseek-v4.1-flash":{"prompt":"0.15","completion":"0.60"}}}]}]}}`
-	if rr := rulesDo(t, g, http.MethodPut, seed); rr.Code != http.StatusOK {
-		t.Fatalf("seed = %d: %s", rr.Code, rr.Body.String())
-	}
-
-	// 1. persisted to the file the process loaded from.
-	raw, err := os.ReadFile(cfg.Path)
-	if err != nil {
-		t.Fatalf("read config: %v", err)
-	}
-	if !strings.Contains(string(raw), "commandcode.ai") {
-		t.Errorf("★ the rule is not in config.yaml on disk — an edit that does not "+
-			"survive a restart:\n%s", raw)
-	}
-	reloaded, err := config.Load(cfg.Path)
-	if err != nil {
-		t.Fatalf("reload saved config: %v", err)
-	}
-	if reloaded.BillingDSL == nil || len(reloaded.BillingDSL.Profiles) == 0 {
-		t.Fatalf("saved config has no billing DSL: %s", raw)
-	}
-	// A round-trip through YAML must not lose the price strings.
-	found := false
-	for _, r := range reloaded.BillingDSL.Profiles[0].Rules {
-		if strings.Contains(r.URL, "commandcode") {
-			found = true
-			if m, ok := r.Models["deepseek-v4.1-flash"]; !ok || m.Prompt != "0.15" {
-				t.Errorf("price did not survive the YAML round-trip: %+v", m)
-			}
-		}
-	}
-	if !found {
-		t.Errorf("the token rule is missing after reload:\n%s", raw)
-	}
-
-	// 2. reached the plugin. Prices live in a separate Lua field from `state`,
-	// and the GET /state endpoint returns the published state together with
-	// the prices sidecar — which is exactly what the plugin's own UI reads.
-	state := g.core.Plugins().State("billing")
-	if state == nil {
-		t.Fatalf("billing plugin published no state")
-	}
-	rr := doReq(t, g, http.MethodGet, "/api/plugins/billing/state", "")
-	if rr.Code != http.StatusOK {
-		t.Fatalf("GET state = %d: %s", rr.Code, rr.Body.String())
-	}
-	var envelope struct {
-		Prices map[string]interface{} `json:"prices"`
-	}
-	if err := json.Unmarshal(rr.Body.Bytes(), &envelope); err != nil {
-		t.Fatalf("decode state envelope: %v", err)
-	}
-	prices := envelope.Prices
-	if prices == nil {
-		t.Fatalf("state payload carries no prices")
-	}
-	sources, ok := prices["sources"].(map[string]interface{})
-	if !ok {
-		t.Fatalf("prices has no sources map: %#v", prices)
-	}
-	cc, ok := sources["commandcode"]
-	if !ok {
-		t.Fatalf("★ the rule was saved but never injected: prices has %v, "+
-			"no commandcode entry. Every request on it stays unpriced.", keysOf(sources))
-	}
-	if cc == nil {
-		t.Error("commandcode price entry is nil")
-	}
-	// The free rule must price localzen, and the two must not be conflated.
-	if _, ok := sources["localzen"]; !ok {
-		t.Errorf("the free rule did not reach the plugin: %v", keysOf(sources))
-	}
-	// The injected price must be per-token, i.e. 0.15 USD/M => 1.5e-7.
-	// Asserting the exact magnitude catches a unit error (per-million vs
-	// per-token), which would be off by a factor of a million and still look
-	// like "a number".
-	models, ok := cc.(map[string]interface{})["models"].(map[string]interface{})
-	if ok {
-		if m, ok := models["deepseek-v4.1-flash"].(map[string]interface{}); ok {
-			p, _ := m["prompt"].(float64)
-			if p <= 0 || p > 1e-6 {
-				t.Errorf("injected prompt price = %v, want per-token (~1.5e-7); "+
-					"a per-million value here would over-bill by 1e6x", p)
-			}
-		}
-	}
-}
-
-func keysOf(m map[string]interface{}) []string {
-	out := make([]string, 0, len(m))
-	for k := range m {
-		out = append(out, k)
-	}
-	sort.Strings(out)
-	return out
-}
-
-// TestBillingRulesRequireAdmin keeps price rewriting behind the admin role,
-// like every other plugin write: prices are configuration, not user data.
-func TestBillingRulesRequireAdmin(t *testing.T) {
-	g := rulesGateway(t)
-	rec, err := g.core.CreateKey("viewer", "user", nil, "")
-	if err != nil {
-		t.Fatal(err)
-	}
-	userKey := rec.Key
-	for _, m := range []string{http.MethodPut, http.MethodPost, http.MethodDelete} {
-		req, _ := http.NewRequest(m, "/api/plugins/billing/rules", strings.NewReader(`{}`))
-		req.Header.Set("Authorization", "Bearer "+userKey)
-		req.Header.Set("Content-Type", "application/json")
-		if rr := newRecorderFor(t, g, req); rr.Code != http.StatusForbidden {
-			t.Errorf("user %s rules = %d, want 403", m, rr.Code)
-		}
-	}
-}

internal/gateway/chat.go

@@ -13,29 +13,87 @@ import (
 	"sync/atomic"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/lua"
+	"modelrouter/internal/lua"
-	"llmsproxy/internal/provider"
+	"modelrouter/internal/provider"
-	"llmsproxy/internal/scheduler"
+	"modelrouter/internal/scheduler"
-	"llmsproxy/internal/types"
+	"modelrouter/internal/types"
 )
 
 // chatRequest mirrors the OpenAI chat completions request the gateway accepts.
+// chatRequest is the gateway's IN-BOUND client shape for /v1/chat/completions.
+//
+// It is deliberately an OpenAI wire struct, NOT the internal unified format:
+// this is one client protocol, and the kernel parses it before anything else
+// knows which protocol the rest of the request will speak. What reaches an
+// adapter is types.UnifiedRequest (the maximal set), which Thinking populates.
 type chatRequest struct {
-	Model           string                 `json:"model"`
+	Model       string              `json:"model"`
-	Messages        []types.ChatMessage    `json:"messages"`
+	Messages    []types.ChatMessage `json:"messages"`
-	Temperature     *float64               `json:"temperature,omitempty"`
+	Temperature *float64            `json:"temperature,omitempty"`
-	MaxTokens       int                    `json:"max_tokens,omitempty"`
+	MaxTokens   int                 `json:"max_tokens,omitempty"`
-	Stream          bool                   `json:"stream,omitempty"`
+	Stream      bool                `json:"stream,omitempty"`
-	Tools           []interface{}          `json:"tools,omitempty"`
+	Tools       []interface{}       `json:"tools,omitempty"`
-	ToolChoice      interface{}            `json:"tool_choice,omitempty"`
+	ToolChoice  interface{}         `json:"tool_choice,omitempty"`
+	// DisableThinking is the OpenAI-ecosystem spelling of "turn thinking off".
+	// It is translated into the protocol-neutral Thinking field on the way in
+	// (see thinkingFromClient), so that flag never travels under this name.
 	DisableThinking bool                   `json:"disable_thinking"`
 	ExtraBody       map[string]interface{} `json:"extra_body,omitempty"`
 
+	// Thinking accepts the neutral shape directly, so a client already speaking
+	// the unified format (or an adapter-served protocol that names it that way)
+	// does not have to round-trip through disable_thinking.
+	Thinking *types.ThinkingConfig `json:"thinking,omitempty"`
+
 	// PromptCacheKey 是 OpenAI 原生的缓存键；部分客户端用它携带会话标识。
 	PromptCacheKey string `json:"prompt_cache_key,omitempty"`
 }
 
+// thinkingFromClient folds the client's spelling of "no thinking" into the
+// neutral Thinking field.
+//
+// Precedence: an explicit `thinking` object wins, because it carries a budget
+// the boolean cannot express. disable_thinking only ever sets Enabled=false —
+// it says nothing about how much, and inventing a budget there would silently
+// override the upstream's default.
+func (r *chatRequest) thinkingFromClient() *types.ThinkingConfig {
+	if r.Thinking != nil {
+		return r.Thinking
+	}
+	if r.DisableThinking {
+		no := false
+		return &types.ThinkingConfig{Enabled: &no}
+	}
+	return nil
+}
+
+// rawTools / rawToolChoice re-encode the loosely-typed client fields into the
+// RawMessage the internal format carries. The kernel never interprets them:
+// they are the client's payload, passed through to the adapter, which is the
+// only layer that knows the tool schema of any particular protocol.
+func (r *chatRequest) rawTools() json.RawMessage {
+	if len(r.Tools) == 0 {
+		return nil
+	}
+	b, err := json.Marshal(r.Tools)
+	if err != nil {
+		return nil
+	}
+	return b
+}
+
+func (r *chatRequest) rawToolChoice() json.RawMessage {
+	if r.ToolChoice == nil {
+		return nil
+	}
+	b, err := json.Marshal(r.ToolChoice)
+	if err != nil {
+		return nil
+	}
+	return b
+}
+
 // ChatCompletion is the non-streaming OpenAI response object.
 type ChatCompletion struct {
 	ID      string            `json:"id"`
@@ -256,17 +314,17 @@ func (g *Gateway) checkQuota(ctx context.Context, model string) *quotaRejection
 		}
 		return nil
 	}
-// A scope that simply does not list the model is a different problem from
+	// A scope that simply does not list the model is a different problem from
-// a quota being exhausted, and the message has to say which: telling an
+	// a quota being exhausted, and the message has to say which: telling an
-// operator "not allowed for this key" when they configured a quota sends
+	// operator "not allowed for this key" when they configured a quota sends
-// them to the wrong setting. AUTO is checked here before the chain is even
+	// them to the wrong setting. AUTO is checked here before the chain is even
-// consulted, so this is the message that surfaces for AUTO.
+	// consulted, so this is the message that surfaces for AUTO.
-if isAuto(model) {
+	if isAuto(model) {
-return &quotaRejection{msg: fmt.Sprintf(
+		return &quotaRejection{msg: fmt.Sprintf(
-"model %q is not in this key's model scope, so it cannot use AUTO; "+
+			"model %q is not in this key's model scope, so it cannot use AUTO; "+
-"add %q to the key's models or remove the scope restriction", model, model)}
+				"add %q to the key's models or remove the scope restriction", model, model)}
-}
+	}
-return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
+	return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
 }
 
 // quotaWindowSuffix describes a quota's reset window for an error message, so
@@ -424,15 +482,15 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 		done := g.stats.Begin()
 		defer done()
 		inner := &types.ChatRequest{
-			Messages:        req.Messages,
+			Messages:      req.Messages,
-			Temperature:     req.Temperature,
+			Temperature:   req.Temperature,
-			MaxTokens:       req.MaxTokens,
+			MaxTokens:     req.MaxTokens,
-			Stream:          req.Stream,
+			Stream:        req.Stream,
-			Tools:           req.Tools,
+			Tools:         req.rawTools(),
-			ToolChoice:      req.ToolChoice,
+			ToolChoice:    req.rawToolChoice(),
-			DisableThinking: req.DisableThinking,
+			Thinking:      req.thinkingFromClient(),
-			ExtraBody:       req.ExtraBody,
+			ExtraBody:     req.ExtraBody,
-			ClientSession:   clientSession,
+			ClientSession: clientSession,
 		}
 		rec := &Req{
 			Key:  keyID(reqKey(ctx)),
@@ -484,16 +542,16 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 	defer done()
 
 	inner := &types.ChatRequest{
-		Model:           effective,
+		Model:         effective,
-		Messages:        req.Messages,
+		Messages:      req.Messages,
-		Temperature:     req.Temperature,
+		Temperature:   req.Temperature,
-		MaxTokens:       req.MaxTokens,
+		MaxTokens:     req.MaxTokens,
-		Stream:          req.Stream,
+		Stream:        req.Stream,
-		Tools:           req.Tools,
+		Tools:         req.rawTools(),
-		ToolChoice:      req.ToolChoice,
+		ToolChoice:    req.rawToolChoice(),
-		DisableThinking: req.DisableThinking,
+		Thinking:      req.thinkingFromClient(),
-		ExtraBody:       req.ExtraBody,
+		ExtraBody:     req.ExtraBody,
-		ClientSession:   clientSession,
+		ClientSession: clientSession,
 	}
 	rec := &Req{
 		Key:    keyID(reqKey(ctx)),
@@ -586,7 +644,7 @@ func estimatePromptTokens(req *types.ChatRequest) int64 {
 	}
 	b, _ := json.Marshal(struct {
 		Messages []types.ChatMessage `json:"messages"`
-		Tools    []interface{}       `json:"tools,omitempty"`
+		Tools    json.RawMessage     `json:"tools,omitempty"`
 	}{Messages: req.Messages, Tools: req.Tools})
 	if len(b) == 0 {
 		return 0
@@ -796,7 +854,7 @@ func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedRespon
 	// keeps the billing plugin from guessing.
 	//
 	// Leaving it unrecorded is worse than useless: the adapters already parse it
-	// and forward it (opencodego/gemini/anthropic all do), so the number exists
+	// and forward it (gemini/anthropic/qwen38 all do), so the number exists
 	// in the response and is thrown away at the one place that could keep it.
 	if d := resp.TokenUsage.CompletionTokensDetails; d != nil && d.ReasoningTokens > 0 {
 		rec.Reasoning = int64(d.ReasoningTokens)
@@ -817,29 +875,21 @@ func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedRespon
 	}
 }
 
-// writeChatCompletion renders a unified response as an OpenAI
+// writeChatCompletion renders a unified response through the out-bound
-// chat.completion object. modelName is the id clients see as the serving
+// Serializer. The default serializer reproduces the OpenAI wire shape this
-// model: the requested id for direct routes, the exact slot model for AUTO.
+// function used to build inline (locked byte-for-byte by e2e golden); an
-func writeChatCompletion(w http.ResponseWriter, resp *types.UnifiedResponse, modelName string) {
+// adapter that declares serialize_response overrides it. modelName is the id
-	msg := RespMessage{Role: "assistant", Content: resp.Content}
+// clients see as the serving model: the requested id for direct routes, the
-	if resp.ReasoningContent != "" {
+// exact slot model for AUTO.
-		msg.ReasoningContent = resp.ReasoningContent
+func writeChatCompletion(w http.ResponseWriter, s Serializer, resp *types.UnifiedResponse, modelName string) {
+	body, err := s.NonStream(resp, modelName, newWireID(), newCreated())
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, "serialize_error", "failed to serialize response")
+		return
 	}
-	if len(resp.ToolCalls) > 0 {
+	w.Header().Set("Content-Type", "application/json")
-		msg.ToolCalls = toolCallsWire(resp.ToolCalls)
+	w.WriteHeader(http.StatusOK)
-	}
+	_, _ = w.Write(body)
-	out := ChatCompletion{
-		ID:      newID(),
-		Object:  "chat.completion",
-		Created: time.Now().Unix(),
-		Model:   modelName,
-		Choices: []ChatChoice{{Index: 0, Message: msg, FinishReason: resp.FinishReason}},
-		Cost:    resp.Cost,
-	}
-	if resp.TokenUsage.Total > 0 || resp.TokenUsage.Prompt > 0 || resp.TokenUsage.Completion > 0 {
-		out.Usage = &resp.TokenUsage
-	}
-	writeJSON(w, http.StatusOK, out)
 }
 
 // singleChat runs a direct (model-pinned) non-streaming request across the
@@ -864,7 +914,7 @@ func (g *Gateway) singleChat(w http.ResponseWriter, ctx context.Context, cands [
 	// the total latency.
 	rec.FirstByteMs = rec.LatMs
 	g.writeRec(rec)
-	writeChatCompletion(w, resp, effective)
+	writeChatCompletion(w, g.serializerFor(usedSrc, cands), resp, effective)
 }
 
 // writeRec records a finished request (audit + aggregates) and fires the
@@ -1059,9 +1109,9 @@ func (g *Gateway) fireEnd(rec *Req) {
 		// A SUBSET of completion_tokens, not an extra. Billing plugins that
 		// add it to completion would double-charge every thinking token.
 		"reasoning_tokens": rec.Reasoning,
-		"image_count":    rec.ImageCount,
+		"image_count":      rec.ImageCount,
-		"error":          rec.Err,
+		"error":            rec.Err,
-		"time":           rec.Time,
+		"time":             rec.Time,
 		// chain_walk: the AUTO tier-by-tier trace, when the request went
 		// through the chain. Empty for a direct request and for a gateway with
 		// no plugins loaded. Absent rather than empty so a plugin can tell
@@ -1116,24 +1166,29 @@ func mergeUsage(prev, cur *types.TokenUsage) *types.TokenUsage {
 	return &out
 }
 
-// pumpStream writes the full SSE sequence for a started stream: role
+// pumpStream writes the full SSE sequence for a started stream. Every byte on
-// preamble, one chunk per unified delta, the terminating finish_reason, the
+// the wire comes from the Serializer, so the OpenAI frame shapes live in
-// OpenAI-standard final usage chunk (empty choices) and [DONE]. modelName
+// serializer.go rather than here, and an adapter that declares serialize_chunk
-// follows writeChatCompletion's rule (requested id for direct routes, exact
+// replaces them wholesale.
-// slot model for AUTO).
+//
-func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan types.UnifiedChunk, modelName string, t0 time.Time) {
+// This function owns only transport concerns — headers, the flush cadence,
+// first-byte timing, and audit accounting — plus the protocol-agnostic rule
+// that usage seen anywhere in the stream becomes the recorded usage. Whether a
+// trailing usage frame is emitted, and in what shape, is the serializer's
+// decision.
+func (g *Gateway) pumpStream(w http.ResponseWriter, ser Serializer, rec *Req, chunks <-chan types.UnifiedChunk, modelName string, t0 time.Time) {
 	w.Header().Set("Content-Type", "text/event-stream")
 	w.Header().Set("Cache-Control", "no-cache")
 	w.Header().Set("Connection", "keep-alive")
 	w.WriteHeader(http.StatusOK)
 
 	flusher, _ := w.(http.Flusher)
-	id := newID()
+	id := newWireID()
-	created := time.Now().Unix()
+	created := newCreated()
-	send := func(obj interface{}) bool {
+	st := &StreamState{}
-		b, err := json.Marshal(obj)
+	send := func(b []byte) bool {
-		if err != nil {
+		if len(b) == 0 {
-			return false
+			return true
 		}
 		if _, err := fmt.Fprintf(w, "data: %s\n\n", b); err != nil {
 			return false
@@ -1144,115 +1199,66 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 		return true
 	}
 
-	if !send(ChatChunk{
+	if b, err := ser.OpenRolePreamble(modelName, id, created, st); err == nil {
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		if !send(b) {
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{Role: "assistant"}}},
+			return
-	}) {
+		}
-		return
 	}
 	// First SSE byte sent to the client: record time-to-first-byte for the
 	// source's status-page latency average.
 	rec.FirstByteMs = time.Since(t0).Milliseconds()
-	var lastUsage *types.TokenUsage
+
-	var lastCost string
-	lastFinish := ""
 	for ck := range chunks {
 		if ck.Usage != nil {
-			lastUsage = mergeUsage(lastUsage, ck.Usage)
+			st.Usage = mergeUsage(st.Usage, ck.Usage)
 		}
 		if ck.Cost != "" {
-			lastCost = ck.Cost
+			st.Cost = ck.Cost
 		}
-		chunk := ChatChunk{
+		b, err := ser.Chunk(&ck, modelName, id, created, st)
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		if err != nil {
+			return
 		}
-		delta := RespMessage{Role: "assistant", Content: ck.Content}
-		if ck.ReasoningContent != "" {
-			delta.ReasoningContent = ck.ReasoningContent
-		}
-		if len(ck.ToolCalls) > 0 {
-			delta.ToolCalls = ck.ToolCalls
-		}
-		choice := ChunkChoice{Index: 0, Delta: delta}
-		if ck.Done {
-			fin := ck.FinishReason
-			if fin == "" {
-				fin = "stop"
-			}
-			lastFinish = fin
-			choice.FinishReason = &fin
-		}
-		chunk.Choices = []ChunkChoice{choice}
 		rec.Compl += int64(len(ck.Content)+len(ck.ReasoningContent)+len(ck.ToolCalls)) / 3
-		if !send(chunk) {
+		if !send(b) {
 			return
 		}
 	}
-	finalFinish := lastFinish
+
-	if finalFinish == "" {
+	// Write the upstream's exact usage back onto the audit record. Without this
-		finalFinish = "stop"
+	// the streamed path kept only the per-chunk byte estimate, so the same
-	}
-	send(ChatChunk{
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &finalFinish}},
-	})
-	// Final usage chunk (OpenAI standard: empty choices + usage before [DONE]).
-	// Prefer the upstream's exact usage if the stream carried it; fall back to
-	// the gateway's estimate otherwise.
-	// Write the upstream's exact numbers back onto the audit record. Without
-	// this the streamed path kept only the per-chunk byte estimate, so the same
 	// request recorded ~1.4x its real completion tokens (measured: upstream 100,
 	// audit 145) while the non-streaming path recorded 100. Two paths, two
 	// different numbers for one request is a reporting bug, not a rounding one.
-	if lastUsage != nil {
+	if st.Usage != nil {
-		if lastUsage.Prompt > 0 {
+		if st.Usage.Prompt > 0 {
-			rec.Prompt = int64(lastUsage.Prompt)
+			rec.Prompt = int64(st.Usage.Prompt)
 		}
-		if lastUsage.Completion > 0 {
+		if st.Usage.Completion > 0 {
-			rec.Compl = int64(lastUsage.Completion)
+			rec.Compl = int64(st.Usage.Completion)
-		}
-	}
-	var tut *types.TokenUsage
-	if lastUsage != nil {
-		tut = lastUsage
-		// Write the upstream's cache accounting back onto the request record
-		// so the audit trail carries the hit/miss split for streaming too.
-		if d := tut.PromptTokensDetails; d != nil {
-			rec.CacheReported = true
-			if d.CachedTokens > 0 {
-				rec.CacheHit = int64(d.CachedTokens)
-			}
-		} else if tut.PromptCacheHit > 0 {
-			rec.CacheHit = int64(tut.PromptCacheHit)
-			rec.CacheReported = true
-		}
-		if tut.PromptCacheMiss > 0 {
-			rec.CacheMiss = int64(tut.PromptCacheMiss)
 		}
+		applyStreamUsageDetails(rec, st.Usage)
 	} else if rec.Prompt+rec.Compl > 0 {
-		u := types.TokenUsage{
+		// No upstream usage on the stream: the recorded per-chunk byte
+		// estimate is the best number available, so hand it to the serializer
+		// as the usage rather than dropping it.
+		st.Usage = &types.TokenUsage{
 			Prompt:     int(rec.Prompt),
 			Completion: int(rec.Compl),
 			Total:      int(rec.Prompt + rec.Compl),
 		}
-		tut = &u
 	}
-	if tut != nil {
+
-		send(ChatChunk{
+	if frames, err := ser.Close(modelName, id, created, st); err == nil {
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		for _, b := range frames {
-			Choices: []ChunkChoice{},
+			if !send(b) {
-			Usage:   tut,
+				return
-			Cost:    lastCost,
+			}
-		})
+		}
-	} else if lastCost != "" {
+	}
-		// Cost arrived without any usage (OpenCode sends it on its own frame).
+	if marker := ser.DoneMarker(); marker != "" {
-		send(ChatChunk{
+		fmt.Fprintf(w, "data: %s\n\n", marker)
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
-			Choices: []ChunkChoice{},
-			Cost:    lastCost,
-		})
 	}
-	fmt.Fprintf(w, "data: [DONE]\n\n")
 	if flusher != nil {
 		flusher.Flush()
 	}
@@ -1283,7 +1289,7 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 	rec.Source = usedSrc
 	g.fireRouted(ctx, "stream", usedSrc, rec.Model, -1, true)
 	rec.Prompt = estimatePromptTokens(req)
-	g.pumpStream(w, rec, chunks, effective, t0)
+	g.pumpStream(w, g.serializerFor(usedSrc, cands), rec, chunks, effective, t0)
 }
 
 // singleChatAuto runs a non-streaming AUTO request down the chain (see
@@ -1316,7 +1322,7 @@ func (g *Gateway) singleChatAuto(w http.ResponseWriter, ctx context.Context, cha
 	g.fireRouted(ctx, "chat", usedSrc, usedModel, -2, false)
 	rec.FirstByteMs = rec.LatMs
 	g.writeRec(rec)
-	writeChatCompletion(w, resp, usedModel)
+	writeChatCompletion(w, g.serializerFor(usedSrc, nil), resp, usedModel)
 }
 
 // streamChatAuto streams an AUTO request down the chain. A slot is abandoned
@@ -1346,7 +1352,7 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 	rec.Source = usedSrc
 	g.fireRouted(ctx, "stream", usedSrc, rec.Model, -2, true)
 	rec.Prompt = estimatePromptTokens(req)
-	g.pumpStream(w, rec, chunks, usedModel, t0)
+	g.pumpStream(w, g.serializerFor(usedSrc, nil), rec, chunks, usedModel, t0)
 }
 
 func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
@@ -1459,3 +1465,39 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 		Data:    resp.ImageData,
 	})
 }
+
+// applyStreamUsageDetails mirrors the upstream's reasoning and cache numbers
+// onto a STREAMED request's audit record.
+//
+// It exists as a named function because this was inline in the streaming handler,
+// where it was untestable — and untested it stayed wrong. The non-streaming
+// recorder read completion_tokens_details.reasoning_tokens; this one did not. So
+// a source's streamed requests recorded completion and no reasoning at all while
+// its non-streamed calls recorded both. Nothing failed and no test was red: the
+// audit trail was simply quieter than it should have been.
+//
+// Found by reading the production audit file after wiring a reasoning-heavy
+// source: completion=8166 with reasoning absent, on every streamed request.
+//
+// Reasoning is a SUBSET of completion, never an addition to it. See
+// recordChatUsage for why that matters to billing.
+func applyStreamUsageDetails(rec *Req, u *types.TokenUsage) {
+	if d := u.CompletionTokensDetails; d != nil && d.ReasoningTokens > 0 {
+		rec.Reasoning = int64(d.ReasoningTokens)
+	}
+	// Cache accounting is here for the same reason it is kept apart: "zero hits
+	// because nothing was cached" and "this provider never reports caching" are
+	// the same number and mean opposite things.
+	if d := u.PromptTokensDetails; d != nil {
+		rec.CacheReported = true
+		if d.CachedTokens > 0 {
+			rec.CacheHit = int64(d.CachedTokens)
+		}
+	} else if u.PromptCacheHit > 0 {
+		rec.CacheHit = int64(u.PromptCacheHit)
+		rec.CacheReported = true
+	}
+	if u.PromptCacheMiss > 0 {
+		rec.CacheMiss = int64(u.PromptCacheMiss)
+	}
+}

internal/gateway/chat_cache_test.go

@@ -3,7 +3,7 @@ package gateway
 import (
 	"testing"
 
-	"llmsproxy/internal/types"
+	"modelrouter/internal/types"
 )
 
 // TestMergeUsageKeepsZeroCacheDetails locks the streaming counterpart of the

internal/gateway/chat_reasoning_test.go

@@ -3,13 +3,13 @@ package gateway
 import (
 	"testing"
 
-	"llmsproxy/internal/types"
+	"modelrouter/internal/types"
 )
 
 // TestRecordChatUsageKeepsReasoningTokens pins the reasoning-token plumbing.
 //
 // The adapters already parse completion_tokens_details.reasoning_tokens and
-// forward it (opencodego, gemini, anthropic all do), and types.go has carried
+// forward it (qwen38, gemini, anthropic all do), and types.go has carried
 // the field since it was defined — but nothing ever READ it. The number arrived
 // in the response and was discarded at the one place that could have kept it,
 // so a pricing rule had no way to split thinking from output.
@@ -69,3 +69,92 @@ func TestRecordChatUsageReasoningAbsentStaysZero(t *testing.T) {
 		}
 	}
 }
+
+// TestStreamedReasoningIsRecordedOnTheRequest is the streaming counterpart.
+//
+// TestRecordChatUsageKeepsReasoningTokens covers the non-streaming path only,
+// and for a long time the streamed path dropped the number: the non-streaming
+// recorder read TokenUsage.CompletionTokensDetails, while the streaming one read
+// only Prompt and Completion from the terminal usage chunk. (Cache accounting was
+// mirrored; reasoning simply never was.)
+//
+// The symptom is invisible in the response — the client sees a correct answer
+// either way. It only shows in the audit trail, where a reasoning-heavy source's
+// streamed requests recorded completion=8166 and no reasoning at all. Found by
+// reading the production audit file, not by a failing test, which is exactly why
+// the property is pinned here.
+func TestStreamedReasoningIsRecordedOnTheRequest(t *testing.T) {
+	rec := &Req{}
+	applyStreamUsageDetails(rec, &types.TokenUsage{
+		Prompt:     159579,
+		Completion: 751,
+		Total:      160330,
+		CompletionTokensDetails: &types.CompletionTokensDetails{
+			ReasoningTokens: 512,
+		},
+	})
+	if rec.Reasoning != 512 {
+		t.Errorf("streamed Reasoning = %d, want 512 — a streamed request records no "+
+			"reasoning, so a pricing rule that splits thinking from output silently "+
+			"bills it as ordinary output", rec.Reasoning)
+	}
+}
+
+// TestStreamedReasoningIsNeverAddedToCompletion pins the SUBSET relation in the
+// streaming path specifically. A presence check would pass if someone wrote
+// rec.Compl += reasoning, and billing would then charge every thinking token
+// twice.
+func TestStreamedReasoningIsNeverAddedToCompletion(t *testing.T) {
+	rec := &Req{}
+	rec.Compl = 751 // set upstream of the call, as the streaming handler does
+	applyStreamUsageDetails(rec, &types.TokenUsage{
+		Completion:              751,
+		CompletionTokensDetails: &types.CompletionTokensDetails{ReasoningTokens: 512},
+	})
+	if rec.Compl != 751 {
+		t.Errorf("Compl = %d, want 751 unchanged — reasoning is a subset, not an addition", rec.Compl)
+	}
+	if rec.Reasoning >= rec.Compl {
+		t.Errorf("Reasoning (%d) must be a SUBSET of Compl (%d)", rec.Reasoning, rec.Compl)
+	}
+}
+
+// TestStreamedUsageWithoutReasoningLeavesItUnset: an upstream that reports no
+// reasoning must not be turned into a reported zero. In the audit trail 0 means
+// "not reported", and a plugin cannot tell that from "no reasoning happened".
+func TestStreamedUsageWithoutReasoningLeavesItUnset(t *testing.T) {
+	rec := &Req{}
+	applyStreamUsageDetails(rec, &types.TokenUsage{Prompt: 10, Completion: 4, Total: 14})
+	if rec.Reasoning != 0 {
+		t.Errorf("Reasoning = %d for an upstream that reported none", rec.Reasoning)
+	}
+}
+
+// TestStreamedCacheAccountingStillWorks is a regression guard for the block the
+// reasoning lines were added to. They share a function, so a future edit that
+// short-circuits one would break the other — silently, since a missing cache
+// number reads as "nothing was cached".
+func TestStreamedCacheAccountingStillWorks(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		u    types.TokenUsage
+		hit  int64
+		miss int64
+		rep  bool
+	}{
+		{"details object", types.TokenUsage{
+			PromptTokensDetails: &types.PromptTokensDetails{CachedTokens: 4096},
+		}, 4096, 0, true},
+		{"legacy hit/miss", types.TokenUsage{
+			PromptCacheHit: 2048, PromptCacheMiss: 512,
+		}, 2048, 512, true},
+		{"nothing reported", types.TokenUsage{}, 0, 0, false},
+	} {
+		rec := &Req{}
+		applyStreamUsageDetails(rec, &c.u)
+		if rec.CacheHit != c.hit || rec.CacheMiss != c.miss || rec.CacheReported != c.rep {
+			t.Errorf("%s: hit=%d miss=%d reported=%v; want %d/%d/%v",
+				c.name, rec.CacheHit, rec.CacheMiss, rec.CacheReported, c.hit, c.miss, c.rep)
+		}
+	}
+}

internal/gateway/deploy_script_test.go

@@ -11,7 +11,7 @@ import (
 
 // deploy.sh is the only release-critical script with no automated coverage, and
 // its newest part — prune_adapter_backups — deletes directories under
-// /etc/llmsproxy. This exercises the function in a throwaway directory with the
+// /etc/modelrouter. This exercises the function in a throwaway directory with the
 // REAL backup-name shape.
 //
 // The shape matters and got this wrong twice: the guard is
@@ -33,34 +33,34 @@ func TestDeployPruneAdapterBackups(t *testing.T) {
 		t.Fatalf("read deploy.sh: %v", err)
 	}
 	src := string(fn)
-	i := strings.Index(src, "prune_adapter_backups()")
+	i := strings.Index(src, "prune_backups()")
 	if i < 0 {
-		t.Fatal("prune_adapter_backups not found in deploy.sh")
+		t.Fatal("prune_backups not found in deploy.sh")
 	}
 	j := strings.Index(src[i:], "\n}\n")
 	if j < 0 {
-		t.Fatal("prune_adapter_backups has no closing brace")
+		t.Fatal("prune_backups has no closing brace")
 	}
 	body := src[i : i+j+3]
 
 	// Redirect ONLY the base path. The logic under test must stay verbatim.
 	body = strings.Replace(body,
-		`local base="/etc/llmsproxy"`,
+		`local base="/etc/modelrouter"`,
 		`local base="`+base+`"`, 1)
-	// KEEP_ADAPTER_BACKUPS must come from deploy.sh itself, NOT from a literal
+	// KEEP_BACKUPS must come from deploy.sh itself, NOT from a literal
 	// here: hardcoding 5 in the shim overrode whatever the script configured, so
-	// a mutation that set KEEP_ADAPTER_BACKUPS=0 in deploy.sh still passed. The
+	// a mutation that set KEEP_BACKUPS=0 in deploy.sh still passed. The
 	// value is part of what is under test.
-	var keepRe = regexp.MustCompile(`(?m)^KEEP_ADAPTER_BACKUPS=(\d+)`)
+	var keepRe = regexp.MustCompile(`(?m)^KEEP_BACKUPS=(\d+)`)
 	keepM := keepRe.FindStringSubmatch(src)
 	if keepM == nil {
-		t.Fatal("deploy.sh no longer sets KEEP_ADAPTER_BACKUPS; the pruning " +
+		t.Fatal("deploy.sh no longer sets KEEP_BACKUPS; the pruning " +
 			"policy would be undefined")
 	}
 	keep := keepM[1]
 
-	script := "warn() { :; }\nlog() { :; }\nKEEP_ADAPTER_BACKUPS=" + keep + "\n" +
+	script := "warn() { :; }\nlog() { :; }\nfail() { echo \"$*\" >&2; exit 1; }\nKEEP_BACKUPS=" + keep + "\n" +
-		body + "\nprune_adapter_backups\n"
+		body + "\nprune_backups adapters\n"
 	scriptPath := filepath.Join(base, "fn.sh")
 	if err := os.WriteFile(scriptPath, []byte(script), 0o600); err != nil {
 		t.Fatalf("write script: %v", err)
@@ -88,7 +88,7 @@ func TestDeployPruneAdapterBackups(t *testing.T) {
 	}
 
 	if out, err := exec.Command("bash", scriptPath).CombinedOutput(); err != nil {
-		t.Fatalf("prune_adapter_backups failed: %v\n%s", err, out)
+		t.Fatalf("prune_backups failed: %v\n%s", err, out)
 	}
 
 	entries, err := os.ReadDir(base)
@@ -102,11 +102,11 @@ func TestDeployPruneAdapterBackups(t *testing.T) {
 		}
 	}
 
-	// 12 well-formed backups with KEEP_ADAPTER_BACKUPS=5 → the newest FIVE
+	// 12 well-formed backups with KEEP_BACKUPS=5 → the newest FIVE
 	// survive (…08 …09 …10 …11 …12) and the older seven are deleted. Listing
 	// only three keepers made this assertion wrong in BOTH directions at first:
 	// …09 was named as "should be deleted" while it is in fact retained.
-	// KEEP_ADAPTER_BACKUPS=5 is a cap on the TOTAL number of backups, and
+	// KEEP_BACKUPS=5 is a cap on the TOTAL number of backups, and
 	// irregular names are never deleted — so they consume slots rather than
 	// being ignored. With 3 unremovable directories present, only 2 of the
 	// well-formed ones can survive. That is the function's actual (and safe)
@@ -202,3 +202,123 @@ func TestDeployBacksUpTheLiveFileBeforeOverwriting(t *testing.T) {
 			"captures the new file and rollback restores nothing")
 	}
 }
+
+// ★ 备份清理必须覆盖 plugins 前缀（真实缺陷）。
+//
+// 事实：backup 清理最初只处理 `adapters.bak.*`，而 `sync_plugins` 同样每次部署
+// 建一个 `plugins.bak.<时间戳>` —— 于是 adapters 的堆积被修好，plugins 继续涨。
+// 实测线上 plugins.bak 积累到 **36 个**、/etc/modelrouter 涨到 192M。
+//
+// 这是「两处只修一处」的老形状。判据钉住：prune_all_backups 必须同时裁两个前缀，
+// 且新增任何 .bak.<时间戳> 前缀时都要在这里补上（否则又是一个泄漏点）。
+func TestDeployPruneCoversBothBackupPrefixes(t *testing.T) {
+	if _, err := os.Stat("../../deploy.sh"); err != nil {
+		t.Skip("deploy.sh not present (packaging-only checkout)")
+	}
+	if _, err := exec.LookPath("bash"); err != nil {
+		t.Skip("bash not available")
+	}
+	fn, err := os.ReadFile("../../deploy.sh")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(fn)
+
+	// ① prune_all_backups 必须存在且两个前缀都裁
+	i := strings.Index(src, "prune_all_backups()")
+	if i < 0 {
+		t.Fatal("prune_all_backups 不存在 —— 调用点若仍只裁 adapters，" +
+			"plugins.bak 会继续无上限堆积（实测 36 个 / 192M）")
+	}
+	j := strings.Index(src[i:], "\n}\n")
+	if j < 0 {
+		t.Fatal("prune_all_backups 没有收尾大括号")
+	}
+	body := src[i : i+j+3]
+	for _, want := range []string{"prune_backups adapters", "prune_backups plugins"} {
+		if !strings.Contains(body, want) {
+			t.Fatalf("prune_all_backups 缺 %q —— plugins.bak 会无上限堆积", want)
+		}
+	}
+
+	// ② ★ 覆盖完整性：脚本里出现的每个 `.bak.<时间戳>` 前缀都必须在裁剪范围内。
+	// 这条比逐字检查两个名字更强 —— 将来有人新增第三种备份目录（比如
+	// runtime.bak.*），这里会立刻报红，而不是等它涨到几百兆。
+	prefixRe := regexp.MustCompile(`([a-z_]+)\.bak\.\$\(date`)
+	found := map[string]bool{}
+	for _, m := range prefixRe.FindAllStringSubmatch(src, -1) {
+		found[m[1]] = true
+	}
+	if len(found) == 0 {
+		t.Fatal("deploy.sh 里没找到任何 .bak.$(date …) 备份目录创建点（判据需更新）")
+	}
+	for prefix := range found {
+		if !strings.Contains(body, "prune_backups "+prefix) {
+			t.Errorf("deploy.sh 会创建 %s.bak.<时间戳> 目录，但 prune_all_backups 不裁它 "+
+				"—— 它就是一个无上限增长的备份泄漏点（实测 plugins.bak 涨到 36 个）", prefix)
+		}
+	}
+
+	// ③ 行为验证：用真函数裁 plugins 前缀，最新 5 份必须留下、更旧的删掉
+	base := t.TempDir()
+	k := strings.Index(src, "prune_backups()")
+	kj := strings.Index(src[k:], "\n}\n")
+	fnBody := src[k : k+kj+3]
+	fnBody = strings.Replace(fnBody, `local base="/etc/modelrouter"`, `local base="`+base+`"`, 1)
+	keepRe := regexp.MustCompile(`(?m)^KEEP_BACKUPS=(\d+)`)
+	km := keepRe.FindStringSubmatch(src)
+	if km == nil {
+		t.Fatal("deploy.sh 不再设置 KEEP_BACKUPS")
+	}
+	script := "warn() { :; }\nlog() { :; }\nfail() { echo \"$*\" >&2; exit 1; }\nKEEP_BACKUPS=" +
+		km[1] + "\n" + fnBody + "\nprune_backups plugins\n"
+	sp := filepath.Join(base, "fn.sh")
+	if err := os.WriteFile(sp, []byte(script), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	for _, ts := range []string{
+		"20260101120000", "20260201120000", "20260301120000", "20260401120000",
+		"20260501120000", "20260601120000", "20260701120000", "20260801120000",
+	} {
+		if err := os.MkdirAll(filepath.Join(base, "plugins.bak."+ts), 0o755); err != nil {
+			t.Fatal(err)
+		}
+	}
+	// adapters 前缀**不该被动**（防止误删别的前缀）
+	if err := os.MkdirAll(filepath.Join(base, "adapters.bak.20260101120000"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if out, err := exec.Command("bash", sp).CombinedOutput(); err != nil {
+		t.Fatalf("prune_backups plugins failed: %v\n%s", err, out)
+	}
+	entries, err := os.ReadDir(base)
+	if err != nil {
+		t.Fatal(err)
+	}
+	got := map[string]bool{}
+	for _, e := range entries {
+		if e.IsDir() {
+			got[e.Name()] = true
+		}
+	}
+	if !got["plugins.bak.20260801120000"] {
+		t.Error("最新的 plugins 备份被删了；裁剪必须从最旧的一端开始")
+	}
+	if got["plugins.bak.20260101120000"] {
+		t.Error("最旧的 plugins 备份存活；裁剪没生效")
+	}
+	if !got["adapters.bak.20260101120000"] {
+		t.Error("裁 plugins 时误删了 adapters 备份 —— 前缀参数必须严格限定范围")
+	}
+
+	// ④ 未知前缀必须响亮失败（防误用：传空串会删掉整个目录）
+	bad := "warn() { :; }\nlog() { :; }\nfail() { echo \"$*\" >&2; exit 1; }\nKEEP_BACKUPS=" +
+		km[1] + "\n" + fnBody + "\nprune_backups \"\"\n"
+	bp := filepath.Join(base, "bad.sh")
+	if err := os.WriteFile(bp, []byte(bad), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if out, err := exec.Command("bash", bp).CombinedOutput(); err == nil {
+		t.Fatalf("prune_backups 接受空前缀（危险：会删整个目录），输出: %s", out)
+	}
+}

internal/gateway/discover.go

@@ -0,0 +1,749 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"strings"
+	"time"
+
+	"modelrouter/internal/config"
+)
+
+// Model discovery: turn "here is a URL and a key" into a configured source.
+//
+// The hard part is not the GET /models call — it is guessing which URL the
+// upstream actually serves. This gateway builds request paths as
+// base_url + adapter endpoint (openai contributes "/chat/completions"), so a
+// base_url that already ends in "/v1" yields "/v1/chat/completions" while a
+// base_url of "/v1" plus our default image path yields the broken
+// "/v1/v1/images/generations". Vendors disagree on whether the version segment
+// belongs to the base or to the path, and there is no header that says which.
+//
+// So discovery does not trust the URL it is given. It tries a small, ordered
+// set of candidates, records which one answered, and reports that back as the
+// normalized base_url + adapter endpoint pair. A caller who typed the wrong
+// form gets a working source anyway, plus the correction needed to make it
+// permanent.
+//
+// What is deliberately NOT here:
+//
+//   - No guessing of model capabilities from the name. "vision" in an id does
+//     not make a model accept images, and marking an image-only endpoint as
+//     kind:"chat" makes every AUTO request fail on it. Kind stays "chat" unless
+//     the upstream's own metadata says otherwise (see discoveredModelKind).
+//   - No live probe per model. A single-token completion per model would cost
+//     real money across a 34-model catalog and takes minutes. /models is the
+//     cheap, authoritative-enough signal for "this id is served here"; whether
+//     it actually answers stays for the normal request path to discover.
+//
+// The one thing that IS verified live is reachability, because a source that
+// 404s is not worth writing to disk.
+
+const (
+	// discoveryTimeout bounds the whole probe. Vendors are routinely slow on
+	// the first request to an uncached endpoint; 20s is enough for a healthy
+	// one and short enough that the UI is not left hanging.
+	discoveryTimeout = 20 * time.Second
+
+	// discoveryModelsPath is the OpenAI-compatible listing endpoint. Nearly
+	// every vendor proxies it verbatim, including the ones whose chat path
+	// differs.
+	discoveryModelsPath = "/v1/models"
+
+	// discoveryMaxBody caps the listing response. A catalog with thousands of
+	// models is still far under this; anything larger is a captive portal or
+	// an error page, not a model list.
+	discoveryMaxBody = 8 << 20
+)
+
+// discoveredModel is one entry of a parsed /models response, kept in the
+// upstream's own terms. The upstream is the only authority on what it serves.
+type discoveredModel struct {
+	ID   string
+	Kind string // "chat" | "image"
+	Meta map[string]interface{}
+}
+
+// discoveryResult reports what a probe found, including the case where nothing
+// worked — the failure text is what the operator needs, so it must name every
+// URL that was tried rather than just the last one.
+type discoveryResult struct {
+	BaseURL  string `json:"base_url"`
+	Endpoint string `json:"endpoint"`
+	// ImageEndpoint is reported separately because the gateway composes it
+	// differently: Provider.ImageURL is BaseURL + ImageEndpoint(), and the
+	// default ImageEndpoint is the ABSOLUTE path "/v1/images/generations".
+	// Chat's default is the relative "/chat/completions". So a base_url that
+	// already ends in /v1 produces a working chat URL and a broken image URL
+	// ("/v1/v1/images/generations") — which is exactly the 404 seen on the
+	// first vela configuration, and which discovery has to correct.
+	ImageEndpoint string   `json:"image_endpoint"`
+	Adapter       string   `json:"adapter"`
+	ModelsPath    string   `json:"models_path"`
+	Models        []string `json:"models"`
+	ImageModels   []string `json:"image_models"`
+	Tried         []string `json:"tried"`
+	LatencyMS     int64    `json:"latency_ms"`
+	Note          string   `json:"note,omitempty"`
+}
+
+// MarshalJSON guarantees the two list fields are always JSON arrays, never
+// null. The WebUI concatenates them directly, and "null" renders as the string
+// "undefined" in every row — a failure that only shows up in a browser and not
+// in a JSON schema check.
+//
+// Relying on the constructor to initialize the slices is not enough: the zero
+// value of this struct is a legitimate thing for a caller to build, and a
+// future code path that forgets the initializer would ship a UI bug silently.
+func (r discoveryResult) MarshalJSON() ([]byte, error) {
+	type alias discoveryResult // avoid recursion
+	out := alias(r)
+	if out.Models == nil {
+		out.Models = []string{}
+	}
+	if out.ImageModels == nil {
+		out.ImageModels = []string{}
+	}
+	if out.Tried == nil {
+		out.Tried = []string{}
+	}
+	return json.Marshal(out)
+}
+
+// discoveryCandidate is one (base_url, models-path) guess. Order matters: the
+// first candidate is the one reported as normalized when several work, because
+// it is the one that keeps base_url free of a version segment and therefore
+// composes correctly with the adapter's chat path.
+type discoveryCandidate struct {
+	baseURL    string
+	modelsPath string
+	// prefer ranks candidates that are more likely to be the vendor's real
+	// path. A 200 from a lower-preference candidate is still accepted, but
+	// only after every higher-preference one failed — see DiscoverModels.
+	prefer int
+}
+
+// discoveryFailure carries the HTTP status the caller should answer with.
+//
+// The status is part of the error rather than left to the handler because the
+// whole point of classifying a failure is that the operator must not have to
+// guess whether the key, the endpoint, or the vendor is at fault.
+type discoveryFailure struct {
+	msg    string
+	status int
+}
+
+func (e *discoveryFailure) Error() string { return e.msg }
+
+// StatusOrDefault reports the intended HTTP status.
+func (e *discoveryFailure) StatusOrDefault() int {
+	if e.status == 0 {
+		return http.StatusBadGateway
+	}
+	return e.status
+}
+
+// discoveryCandidates enumerates the URL forms worth trying for a raw input.
+//
+// The ranking encodes a real observation rather than a guess: the amr-link
+// endpoint serves its catalog at "/v1/models" (the form an SDK example shows)
+// from a base the operator pasted WITH the "/v1", while several other vendors
+// mount it at the host root. Both are probed; the one that answers is reported
+// back as a normalized base/endpoint pair.
+//
+// What we are really deciding is the base_url to persist. The safest one keeps
+// the version segment on the ENDPOINT, because openai's adapter endpoint is
+// "/chat/completions" and the two are concatenated: base "https://h" +
+// endpoint "/v1/chat/completions" is correct, while leaving "/v1" on the base
+// doubles the default image path into "/v1/v1/images/generations" — which is
+// exactly the 404 observed on the first vela configuration attempt.
+func discoveryCandidates(raw string) []discoveryCandidate {
+	raw = strings.TrimRight(strings.TrimSpace(raw), "/")
+	if raw == "" {
+		return nil
+	}
+	// Deduplicate on the URL that will actually be requested, not on the
+	// (base, path) pair. base "https://h/v1" + path "/models" and base "https://h"
+	// + path "/v1/models" both request the same address, and listing it twice in
+	// an error message makes the operator think two things were tried.
+	seenURL := map[string]bool{}
+	seenPair := map[string]bool{}
+	var out []discoveryCandidate
+	add := func(base, path string, prefer int) {
+		pair := base + "|" + path
+		if seenPair[pair] || seenURL[base+path] {
+			return
+		}
+		seenPair[pair] = true
+		seenURL[base+path] = true
+		out = append(out, discoveryCandidate{baseURL: base, modelsPath: path, prefer: prefer})
+	}
+
+	stripped := trimVersionSuffix(raw)
+	hasVersion := stripped != raw
+
+	// Both catalog paths are always tried, whichever form the operator typed.
+	//
+	// The earlier version offered "/models" only when the base already carried a
+	// version segment. That left a bare host ("https://api.vendor.com") with a
+	// SINGLE candidate and therefore no fallback at all — so a vendor serving
+	// its catalog at the bare "/models" was reported as unreachable. Every
+	// test used a "/v1"-suffixed base and so never saw it; siliconflow's
+	// "/v1/models" happens to answer, which is why the gap survived review.
+	//
+	// Preference only decides WHICH URL is reported when several work, never
+	// which ones are attempted.
+	//
+	// Order matters: a base that already ends in a version segment must be
+	// paired with the RELATIVE "/models" first, because pairing it with
+	// "/v1/models" produces "…/v1/v1/models" — a URL that is 404 by construction.
+	// Trying it first would spend a round trip on a certain miss and, worse,
+	// make the reported "tried" list lead with nonsense.
+	if hasVersion {
+		add(raw, "/models", 0)
+		add(raw, discoveryModelsPath, 1) // the doubled form; kept only as a last resort
+		add(stripped, "/models", 2)
+		add(stripped, discoveryModelsPath, 3)
+	} else {
+		add(raw, discoveryModelsPath, 0)
+		add(raw, "/models", 1)
+	}
+
+	// Stable insertion sort by preference; add() above is not ordered.
+	for i := 1; i < len(out); i++ {
+		for j := i; j > 0 && out[j].prefer < out[j-1].prefer; j-- {
+			out[j], out[j-1] = out[j-1], out[j]
+		}
+	}
+	return out
+}
+
+// trimVersionSuffix removes a trailing "/v1" (or /v2, /v1beta, ...) so the
+// host root can be probed on its own.
+func trimVersionSuffix(base string) string {
+	i := strings.LastIndex(base, "/")
+	if i < 0 {
+		return base
+	}
+	if !isVersionSegment(base[i+1:]) {
+		return base
+	}
+	return base[:i]
+}
+
+// versionSegment returns the leading path segment of a models path when that
+// segment is a version marker ("/v1/models" → "/v1", "/models" → ""). It is the
+// piece that must move between base_url and endpoint so the two compose.
+func versionSegment(modelsPath string) string {
+	parts := strings.Split(strings.Trim(modelsPath, "/"), "/")
+	if len(parts) != 2 || !isVersionSegment(parts[0]) {
+		return ""
+	}
+	return "/" + parts[0]
+}
+
+// isVersionSegment reports whether a path segment looks like "v1", "v2",
+// "v1beta" — and not like a real path name such as "vision" or a wildcard
+// path such as "v1x".
+//
+// The rule is: a leading "v", then digits, then an OPTIONAL suffix drawn from
+// the prerelease markers vendors actually use. A free-form alphabetic tail was
+// tried and rejected — it classifies "v1x" as a version, and that is how a
+// templated path segment would get stripped off a real base URL.
+func isVersionSegment(seg string) bool {
+	if len(seg) < 2 || seg[0] != 'v' {
+		return false
+	}
+	i := 1
+	digits := 0
+	for ; i < len(seg); i++ {
+		if seg[i] < '0' || seg[i] > '9' {
+			break
+		}
+		digits++
+	}
+	if digits == 0 {
+		return false
+	}
+	suffix := strings.ToLower(seg[i:])
+	if suffix == "" {
+		return true
+	}
+	for _, marker := range []string{"alpha", "beta", "preview", "dev", "rc"} {
+		if suffix == marker || strings.HasPrefix(suffix, "-"+marker) {
+			return true
+		}
+	}
+	return false
+}
+
+// composeEndpoint derives the chat endpoint to persist, given the models path
+// that actually answered, the base_url being persisted, and the base the
+// operator typed.
+//
+// The rule is a single invariant: exactly one version segment must end up in
+// base_url + endpoint. The adapter contributes "/chat/completions" and the
+// default image path contributes "/v1/images/generations", so a version left on
+// BOTH breaks one of them — and one left on NEITHER sends chat to the wrong
+// path on vendors that require the segment.
+func composeEndpoint(modelsPath, persistedBase, typedBase string) string {
+	// The version segment currently present on the persisted base.
+	onBase := ""
+	if b := trimVersionSuffix(persistedBase); b != persistedBase {
+		onBase = persistedBase[len(b):]
+	}
+	// The version segment present in the models path that answered.
+	inPath := versionSegment(modelsPath)
+
+	// The version segment actually WORKED: prefer the models path's, then the
+	// one the operator typed.
+	version := inPath
+	if version == "" && onBase == "" {
+		if b := trimVersionSuffix(typedBase); b != typedBase && typedBase != "" {
+			version = typedBase[len(b):]
+		}
+	}
+	if version == "" {
+		return "" // adapter default already composes
+	}
+	return version + "/chat/completions"
+}
+
+// composeImageEndpoint derives the image endpoint to persist.
+//
+// Provider.ImageURL is `base_url + ImageEndpoint()`, and the built-in default
+// is the absolute path "/v1/images/generations" — unlike chat's relative
+// "/chat/completions". So the two defaults do NOT agree about who owns the
+// version segment, which is the trap this function exists to close:
+//
+//	base "https://h"     + "/v1/images/generations"  → correct
+//	base "https://h/v1"  + "/v1/images/generations"  → "/v1/v1/…" 404
+//	base "https://h/v1"  + "/images/generations"     → "/v1/images/generations"  correct
+//
+// The second row is not a theoretical guess. Measured against the real
+// amr-link endpoint with base_url "https://amr-link.open-design.ai/v1":
+//
+//	…/v1/images/generations   → 200 (the image is then generated; that call
+//	                              takes minutes, which is the upstream's own
+//	                              latency and not a routing problem)
+//	…/v1/v1/images/generations → 404
+//	…/images/generations      → 404
+//
+// Leaving the default in place when the base already ends in a version is
+// therefore always wrong for such an upstream, and the version-relative form
+// is the one that composes.
+func composeImageEndpoint(persistedBase string) string {
+	if b := trimVersionSuffix(persistedBase); b != persistedBase {
+		return "/images/generations"
+	}
+	return ""
+}
+
+// DiscoverModels probes rawURL for a model catalog. apiKey is sent as a bearer
+// token; some gateways accept an unauthenticated listing, so an empty key is
+// allowed and simply omits the header.
+//
+// It returns on the first candidate that yields a parseable catalog, so the
+// reported latency is that of the working path, not the sum of the failures.
+func DiscoverModels(ctx context.Context, rawURL, apiKey string) (*discoveryResult, error) {
+	cands := discoveryCandidates(rawURL)
+	if len(cands) == 0 {
+		return nil, fmt.Errorf("base url required")
+	}
+
+	client := &http.Client{Timeout: discoveryTimeout}
+	var tried []string
+	var firstErr string
+	var authErr string
+	typedBase := strings.TrimRight(strings.TrimSpace(rawURL), "/")
+
+	var lastErr string
+	for _, c := range cands {
+		full := c.baseURL + c.modelsPath
+		tried = append(tried, full)
+
+		req, err := http.NewRequestWithContext(ctx, http.MethodGet, full, nil)
+		if err != nil {
+			firstErr = err.Error()
+			continue
+		}
+		req.Header.Set("Accept", "application/json")
+		if apiKey != "" {
+			req.Header.Set("Authorization", "Bearer "+apiKey)
+		}
+
+		start := time.Now()
+		resp, err := client.Do(req)
+		if err != nil {
+			lastErr = classifyTransport(err)
+			if lastErr == "" {
+				lastErr = err.Error()
+			}
+			continue
+		}
+		body, readErr := io.ReadAll(io.LimitReader(resp.Body, discoveryMaxBody))
+		resp.Body.Close()
+		took := time.Since(start)
+		if readErr != nil {
+			firstErr = readErr.Error()
+			continue
+		}
+		if resp.StatusCode != http.StatusOK {
+			// An auth rejection is the operator's key being wrong, not the
+			// upstream being down. Keep it: it is the single most common
+			// failure and 502 would send them to the wrong place.
+			if af := classifyAuthStatus(resp.StatusCode, body); af != "" {
+				authErr = af
+				continue
+			}
+			firstErr = fmt.Sprintf("%s returned %d", full, resp.StatusCode)
+			continue
+		}
+
+		models, err := parseModelCatalog(body)
+		if err != nil {
+			firstErr = fmt.Sprintf("%s: %v", full, err)
+			continue
+		}
+
+		res := &discoveryResult{
+			BaseURL:     c.baseURL,
+			ModelsPath:  c.modelsPath,
+			Adapter:     "openai",
+			Models:      []string{},
+			ImageModels: []string{},
+			Tried:       append([]string{}, tried...),
+			LatencyMS:   took.Milliseconds(),
+		}
+		// Persist a base/endpoint pair that composes correctly. The adapter's
+		// own endpoint is "/chat/completions", so any version segment has to be
+		// carried on the endpoint rather than left on the base (where it doubles
+		// the default image path into "/v1/v1/images/generations" — the exact
+		// 404 seen on the first vela configuration attempt).
+		res.Endpoint = composeEndpoint(c.modelsPath, c.baseURL, typedBase)
+		res.ImageEndpoint = composeImageEndpoint(c.baseURL)
+
+		// If the operator typed a version segment and we are persisting a base
+		// that no longer has it, the endpoint above is what restores it. If the
+		// base kept the segment, the adapter default is already correct.
+
+		seen := map[string]bool{}
+		for _, m := range models {
+			if m.ID == "" || seen[m.ID] {
+				continue
+			}
+			seen[m.ID] = true
+			if m.Kind == "image" {
+				res.ImageModels = append(res.ImageModels, m.ID)
+			} else {
+				res.Models = append(res.Models, m.ID)
+			}
+		}
+		sort.Strings(res.Models)
+		sort.Strings(res.ImageModels)
+		if len(res.Models) == 0 && len(res.ImageModels) == 0 {
+			firstErr = fmt.Sprintf("%s: catalog contained no usable model ids", full)
+			continue
+		}
+		if c.baseURL != typedBase {
+			res.Note = fmt.Sprintf("base url normalized from %s to %s; endpoint set to %q so it composes with the adapter",
+				typedBase, c.baseURL, res.Endpoint)
+		}
+		return res, nil
+	}
+	return nil, buildDiscoveryError(tried, authErr, firstErr, lastErr)
+}
+
+// classifyAuthStatus turns a 401/403 from the catalog endpoint into a message
+// that names the credential as the cause.
+//
+// This distinction is the whole point: a wrong key and an unreachable vendor
+// are the two most common discovery failures, and they need opposite actions
+// (re-type the key vs. go check the vendor). Reporting both as "upstream
+// error" makes the operator guess.
+func classifyAuthStatus(status int, body []byte) string {
+	if status != http.StatusUnauthorized && status != http.StatusForbidden {
+		return ""
+	}
+	snippet := strings.TrimSpace(string(body))
+	// Vendors put the reason in wildly different shapes; a compact quote is
+	// more useful than a parse attempt that fails on most of them.
+	snippet = strings.Join(strings.Fields(snippet), " ")
+	if len(snippet) > 200 {
+		snippet = snippet[:200] + "…"
+	}
+	return fmt.Sprintf("upstream rejected the api key (HTTP %d): %s — "+
+		"check the key, not the endpoint", status, snippet)
+}
+
+// classifyTransport turns low-level network failures into something an
+// operator can act on. An empty result means "no specific advice", and the
+// caller keeps the raw error.
+func classifyTransport(err error) string {
+	if err == nil {
+		return ""
+	}
+	msg := err.Error()
+	lower := strings.ToLower(msg)
+	switch {
+	case strings.Contains(lower, "no such host"):
+		return "hostname does not resolve: " + msg
+	case strings.Contains(lower, "certificate"), strings.Contains(lower, "tls"):
+		return "TLS verification failed: " + msg
+	case strings.Contains(lower, "connection refused"):
+		return "connection refused: " + msg
+	}
+	return ""
+}
+
+// buildDiscoveryError assembles the final error, preferring the most specific
+// diagnosis available. Every URL tried is listed: an operator staring at a
+// failed probe needs to know the search actually happened, not just the last
+// attempt.
+func buildDiscoveryError(tried []string, authErr, firstErr, lastErr string) error {
+	triedPart := "no model catalog found"
+	if len(tried) > 0 {
+		triedPart = "no model catalog found; tried " + strings.Join(tried, ", ")
+	}
+	if authErr != "" {
+		return &discoveryFailure{status: http.StatusUnauthorized, msg: authErr + " (" + triedPart + ")"}
+	}
+	detail := firstErr
+	if detail == "" {
+		detail = lastErr
+	}
+	if detail == "" {
+		return errors.New(triedPart)
+	}
+	if lastErr != "" && lastErr != firstErr {
+		detail += " (last: " + lastErr + ")"
+	}
+	return &discoveryFailure{status: http.StatusBadGateway, msg: triedPart + "; " + detail}
+}
+
+// upstreamModel is the subset of a /models entry this code reads. Vendors add
+// fields freely and rename the nesting; only id is universal, and the modality
+// hints below are read opportunistically.
+type modelCaps struct {
+	ImageGeneration bool `json:"image_generation"`
+	Vision          bool `json:"vision"`
+}
+
+type upstreamModel struct {
+	ID           string `json:"id"`
+	Object       string `json:"object"`
+	Kind         string `json:"kind"`
+	Type         string `json:"type"`
+	Modality     string `json:"modality"`
+	Architecture struct {
+		InputModalities  []string `json:"input_modalities"`
+		OutputModalities []string `json:"output_modalities"`
+	} `json:"architecture"`
+	Capabilities modelCaps `json:"capabilities"`
+
+	// Metadata is kept as raw JSON on purpose. The real amr-link catalog nests
+	// cost/capability/context_limit here, and a typed struct would silently
+	// discard every key this build does not know about — the operator would
+	// then lose the context limit their vendor published. Raw also lets
+	// input_modalities/output_modalities be read without a second type.
+	Metadata json.RawMessage `json:"metadata"`
+}
+
+// metadataFields lists the keys worth surfacing in the UI. They are copied
+// through verbatim and nothing is inferred from them; an upstream that sends
+// no context limit must leave that absent rather than have this code write 0.
+var metadataFields = []string{
+	"context_length", "context_window", "max_context_length",
+	"context_limit", "context_budget", "cost", "capability",
+}
+
+// parseModelCatalog reads the OpenAI list shape {"data":[...]} and the bare
+// array shape some gateways return. A response that is valid JSON but has
+// neither shape is an error, not an empty catalog: reporting "0 models" for a
+// login page or a rate-limit JSON would silently create an empty source.
+func parseModelCatalog(body []byte) ([]discoveredModel, error) {
+	// "null" unmarshals into a nil slice without error, and an HTML login page
+	// fails to unmarshal at all. Both must be errors.
+	trimmed := strings.TrimSpace(string(body))
+	if trimmed == "" || trimmed == "null" {
+		return nil, fmt.Errorf("empty response")
+	}
+
+	var wrapper struct {
+		Data   []upstreamModel `json:"data"`
+		Models []upstreamModel `json:"models"`
+	}
+	if err := json.Unmarshal(body, &wrapper); err == nil {
+		if len(wrapper.Data) > 0 {
+			return toDiscovered(wrapper.Data), nil
+		}
+		if len(wrapper.Models) > 0 {
+			return toDiscovered(wrapper.Models), nil
+		}
+		// A well-formed but empty catalog is a legitimate answer from a key
+		// that has no models provisioned. Distinguish it from "not a catalog"
+		// by checking the shape actually parsed.
+		var probe map[string]json.RawMessage
+		if json.Unmarshal(body, &probe) == nil {
+			if _, ok := probe["data"]; ok {
+				return nil, nil
+			}
+			if _, ok := probe["models"]; ok {
+				return nil, nil
+			}
+		}
+	}
+	var bare []upstreamModel
+	if err := json.Unmarshal(body, &bare); err == nil {
+		return toDiscovered(bare), nil
+	}
+	return nil, fmt.Errorf("response is not a model catalog")
+}
+
+func toDiscovered(in []upstreamModel) []discoveredModel {
+	out := make([]discoveredModel, 0, len(in))
+	for _, m := range in {
+		out = append(out, discoveredModel{
+			ID:   m.ID,
+			Kind: discoveredModelKind(m),
+			Meta: modelMeta(m),
+		})
+	}
+	return out
+}
+
+// discoveredModelKind trusts the upstream's own modality fields and nothing
+// else. It never infers from the model NAME: ids like "gpt-image-1" or
+// "ideogram-v2" appear on catalogs that also serve chat, and a wrong kind
+// poisons the AUTO chain (an image-only id selected for a chat request fails
+// every tier behind it).
+func discoveredModelKind(m upstreamModel) string {
+	modIn := append(append([]string{}, m.Architecture.InputModalities...),
+		modalityList(m.Metadata, "input_modalities")...)
+	modOut := append(append([]string{}, m.Architecture.OutputModalities...),
+		modalityList(m.Metadata, "output_modalities")...)
+
+	has := func(list []string, want string) bool {
+		for _, v := range list {
+			if strings.EqualFold(strings.TrimSpace(v), want) {
+				return true
+			}
+		}
+		return false
+	}
+	if m.Capabilities.ImageGeneration || has(modOut, "image") {
+		return "image"
+	}
+	if has(modIn, "image") || has(modIn, "file") || has(modIn, "audio") || m.Capabilities.Vision {
+		return "chat" // multimodal input still means a chat model
+	}
+	// Some catalogs encode modality as a top-level string instead of a list.
+	// Only the image-ish spellings count: "chat" here means the upstream said
+	// chat, and treating that as "unknown" would ignore its statement.
+	switch strings.ToLower(strings.TrimSpace(m.Kind + " " + m.Modality + " " + m.Type)) {
+	case "image", "text-to-image", "image_generation", "text_to_image":
+		return "image"
+	}
+	return "chat"
+}
+
+// modalityList reads one array-of-strings field out of a raw JSON object,
+// tolerating both absence and a non-array value.
+func modalityList(raw json.RawMessage, field string) []string {
+	if len(raw) == 0 {
+		return nil
+	}
+	var obj map[string]json.RawMessage
+	if err := json.Unmarshal(raw, &obj); err != nil {
+		return nil
+	}
+	v, ok := obj[field]
+	if !ok {
+		return nil
+	}
+	var out []string
+	if err := json.Unmarshal(v, &out); err != nil {
+		return nil
+	}
+	return out
+}
+
+// modelMeta carries through the fields an operator wants to see in the UI
+// without this code inventing them. The upstream nests most of them under
+// "metadata" (the real amr-link catalog does) while some put the same keys at
+// the top level, so both are read and top level wins on a tie. Only keys the
+// upstream actually sent are copied, so an absent context limit stays absent
+// instead of becoming 0.
+func modelMeta(m upstreamModel) map[string]interface{} {
+	meta := map[string]interface{}{}
+
+	copyFields := func(raw json.RawMessage) {
+		if len(raw) == 0 {
+			return
+		}
+		var obj map[string]json.RawMessage
+		if err := json.Unmarshal(raw, &obj); err != nil {
+			return
+		}
+		for _, k := range metadataFields {
+			if v, ok := obj[k]; ok {
+				if dec := rawOrString(v); dec != nil {
+					meta[k] = dec
+				}
+			}
+		}
+	}
+
+	copyFields(m.Metadata)  // nested first
+	copyFields(mustJSON(m)) // top level wins
+	if len(meta) == 0 {
+		return nil
+	}
+	return meta
+}
+
+// rawOrString keeps numeric metadata as a number and everything else as its
+// decoded string, so the UI can sort on a cost or context length instead of
+// receiving "1048000" and having to coerce it.
+func rawOrString(raw json.RawMessage) interface{} {
+	var num json.Number
+	if err := json.Unmarshal(raw, &num); err == nil {
+		if f, err := num.Float64(); err == nil {
+			return f
+		}
+	}
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return s
+	}
+	var any interface{}
+	if err := json.Unmarshal(raw, &any); err == nil {
+		return any
+	}
+	return nil
+}
+
+func mustJSON(v interface{}) []byte {
+	b, _ := json.Marshal(v)
+	return b
+}
+
+// modelsToConfig converts discovery output into config.Model entries, applying
+// the same priority convention the UI uses: chat models the operator is
+// likely to want first get a lower number, and image models sort last because
+// they are unreachable from /v1/chat/completions.
+func modelsToConfig(chat, image []string) []config.Model {
+	out := make([]config.Model, 0, len(chat)+len(image))
+	for _, id := range chat {
+		out = append(out, config.Model{ID: id, Kind: "chat"})
+	}
+	for _, id := range image {
+		out = append(out, config.Model{ID: id, Kind: "image"})
+	}
+	return out
+}

internal/gateway/discover_api.go

@@ -0,0 +1,497 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"modelrouter/internal/config"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// HTTP surface for model discovery.
+//
+// Two endpoints, deliberately separate:
+//
+//	POST /api/sources/discover        — probe a URL+key, return the catalog.
+//	                                   Nothing is written.
+//	POST /api/sources/discover/apply  — probe, then write the source.
+//
+// The split exists because discovery mutates nothing and is therefore safe to
+// call repeatedly while the operator is still typing a key, whereas apply
+// rewrites credentials and the AUTO chain's inputs. Making them one endpoint
+// with a flag would mean a "just check it" click can silently create a source.
+//
+// Both are admin-only, matching /api/sources: the request body carries an API
+// key, and a model catalog reveals which upstreams an operator uses.
+
+// discoverRequest is the body of both endpoints.
+type discoverRequest struct {
+	// BaseURL is what the operator pasted. It may or may not include a
+	// version segment; discovery works that out rather than making the
+	// caller know this gateway's path-composition rules.
+	BaseURL string `json:"base_url"`
+	APIKey  string `json:"api_key"`
+
+	// Name is required for apply, ignored for discover. It defaults from the
+	// base URL's host when omitted, because "https://api.vendor.com/v1" is a
+	// better starting name than "new-source".
+	Name string `json:"name"`
+
+	// Adapter is optional. Discovery assumes "openai" because that is what a
+	// /v1/models catalog implies; a non-OpenAI upstream (Anthropic, Gemini)
+	// has no such endpoint, so discovery will legitimately fail for those and
+	// the operator configures them by hand.
+	Adapter string `json:"adapter"`
+
+	// Temperature and MaxConcurrent are carried into the written source.
+	Temperature   float64 `json:"temperature"`
+	MaxConcurrent int     `json:"max_concurrent"`
+
+	// ReplaceModels drops the source's existing model list before applying the
+	// discovered one. It defaults to true for a NEW source and is required to
+	// be spelled out for an existing one, so a re-discovery cannot silently
+	// delete hand-tuned priorities.
+	ReplaceModels *bool `json:"replace_models,omitempty"`
+
+	// Select narrows what gets written. Empty means everything discovered.
+	// Ids that are not in the catalog are an error rather than a silent skip:
+	// a typo here would produce a source that 404s on every request.
+	Include []string `json:"include,omitempty"`
+	Exclude []string `json:"exclude,omitempty"`
+
+	// ImageModels names ids that are **kind=image** but whose upstream catalog
+	// does not say so.
+	//
+	// 为什么需要（真实缺口）：discoveredModelKind 只看上游的 modality 字段，
+	// 并且刻意**不从模型名推断**（那会把 "gpt-image-1" 这类同名聊天模型
+	// 错归成生图，而错误的 kind 会毒死 AUTO 链）。
+	// 但有些网关的 `/v1/models` 只返回 `{id,object,created,owned_by}` ——
+	// 零 modality（实测：某聚合网关的 3 个 gpt-image-* 全部落成 chat，
+	// 于是生图端点永远报 `no image source configured`）。
+	//
+	// 两条原则都不破：不猜（默认仍完全按元数据），但允许**显式声明**。
+	// 只影响写入的 kind，不影响目录里有没有这个模型。
+	ImageModels []string `json:"image_models,omitempty"`
+}
+
+// discoverResponse is the JSON the WebUI binds to.
+type discoverResponse struct {
+	OK bool `json:"ok"`
+
+	// Normalized source fields, ready to be echoed back to /api/sources.
+	BaseURL     string   `json:"base_url"`
+	Adapter     string   `json:"adapter"`
+	Models      []string `json:"models"`
+	ImageModels []string `json:"image_models"`
+
+	// Endpoint is set only when discovery had to move a version segment.
+	Endpoint string `json:"endpoint"`
+	// ImageEndpoint is required whenever base_url carries a version segment:
+	// the gateway's default image path is the ABSOLUTE "/v1/images/generations"
+	// while chat's default is the relative "/chat/completions", so the same
+	// base_url that yields a working chat URL yields a 404 image URL.
+	ImageEndpoint string `json:"image_endpoint"`
+
+	// Diagnostics for the operator.
+	ModelsPath string   `json:"models_path"`
+	Tried      []string `json:"tried"`
+	LatencyMS  int64    `json:"latency_ms"`
+	Note       string   `json:"note,omitempty"`
+	Warnings   []string `json:"warnings,omitempty"`
+
+	// Set by apply only.
+	SourceName string `json:"source_name,omitempty"`
+	Applied    bool   `json:"applied,omitempty"`
+}
+
+func (g *Gateway) handleSourceDiscoverAPI(w http.ResponseWriter, r *http.Request) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	apply := strings.HasSuffix(r.URL.Path, "/apply")
+	if r.Method != http.MethodPost {
+		w.Header().Set("Allow", "POST")
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed",
+			"use POST with {\"base_url\":…,\"api_key\":…}")
+		return
+	}
+
+	body, err := io.ReadAll(io.LimitReader(r.Body, 1<<20))
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "cannot read body")
+		return
+	}
+	var req discoverRequest
+	if err := json.Unmarshal(body, &req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	req.BaseURL = strings.TrimSpace(req.BaseURL)
+	req.Name = strings.TrimSpace(req.Name)
+
+	// ★ 只给 `name`（或两者都给但 key 为空）⇒ **从已存源取 base_url 与凭据**。
+	//
+	// 为什么必须有这条路径（真实痛点）：刷新一个**已存在**源的模型目录，此前
+	// 必须让运维把明文 key 再粘一次 —— 而库里的 key 是 `enc:v1:` 密文，
+	// 运维手里未必还有原文；WebUI 也不回显明文（这是正确的安全设计）。
+	// 结果就是「想刷新目录 → 必须重新申请/粘贴一次 key」，很多人干脆不刷新，
+	// 于是源的模型目录越放越旧（实例：某源的 3 个生图模型从未被登记）。
+	//
+	// 安全边界：① 已要求 admin 角色；② 只能探测**本网关已配置的**源，
+	// 不能借此拿任意 URL + 任意 key 做探测（那样就成了 SSRF/密钥外带渠道）；
+	// ③ 响应不回显 key（只回目录）。
+	if req.Name != "" {
+		if old := g.sourceByName(req.Name); old != nil {
+			if req.BaseURL == "" {
+				req.BaseURL = old.BaseURL
+			}
+			if req.APIKey == "" {
+				req.APIKey = old.APIKey
+			}
+			if req.Adapter == "" {
+				req.Adapter = old.Adapter
+			}
+		}
+	}
+
+	if req.BaseURL == "" {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			"base_url required (or pass name of an existing source to reuse its base_url and credential)")
+		return
+	}
+	if req.Adapter == "" {
+		req.Adapter = "openai"
+	}
+
+	// Bounded so a hung upstream cannot pin the handler; discovery has its own
+	// per-request timeout on top.
+	ctx, cancel := context.WithTimeout(r.Context(), 60*time.Second)
+	defer cancel()
+
+	res, err := DiscoverModels(ctx, req.BaseURL, req.APIKey)
+	if err != nil {
+		// A rejected key must not be reported as an upstream outage: the two
+		// need opposite actions from the operator, and 502 sends them to the
+		// wrong one.
+		status := http.StatusBadGateway
+		var f *discoveryFailure
+		if errors.As(err, &f) {
+			status = f.StatusOrDefault()
+		}
+		writeError(w, status, "discover_failed", err.Error())
+		return
+	}
+
+	// ★ 显式生图声明：把 operator 点名的 id 从 chat 挪到 image。
+	//
+	// 放在 selectModels **之前**：后者按 include/exclude 裁剪两个列表，
+	// 挪类必须在裁剪前完成，否则一个"只声明成生图、又被 include 选中"的
+	// 模型会先在 chat 里被保留、再被这里挪走 —— 顺序反了会得到两份。
+	//
+	// 只认**目录里确实存在的** id：写错一个字不该静默造出一个不存在的模型
+	// （那会让源在每次照片请求时 404），而是明确报错。
+	if len(req.ImageModels) > 0 {
+		have := map[string]bool{}
+		for _, m := range res.Models {
+			have[m] = true
+		}
+		for _, m := range res.ImageModels {
+			have[m] = true
+		}
+		var missing []string
+		force := map[string]bool{}
+		for _, id := range req.ImageModels {
+			if !have[id] {
+				missing = append(missing, id)
+				continue
+			}
+			force[id] = true
+		}
+		if len(missing) > 0 {
+			writeError(w, http.StatusBadRequest, "invalid_request",
+				"image_models contains ids that are not in the upstream catalog: "+
+					strings.Join(missing, ", "))
+			return
+		}
+		var keepChat []string
+		for _, m := range res.Models {
+			if !force[m] {
+				keepChat = append(keepChat, m)
+			}
+		}
+		seenImg := map[string]bool{}
+		var newImage []string
+		for _, m := range res.ImageModels {
+			seenImg[m] = true
+			newImage = append(newImage, m)
+		}
+		for id := range force {
+			if !seenImg[id] {
+				newImage = append(newImage, id)
+			}
+		}
+		res.Models = keepChat
+		res.ImageModels = newImage
+	}
+
+	chat, image, warns, err := selectModels(res.Models, res.ImageModels, req.Include, req.Exclude)
+	if err != nil {
+		// 400: the operator asked for a model the catalog does not have.
+		writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+		return
+	}
+
+	out := &discoverResponse{
+		OK:            true,
+		BaseURL:       res.BaseURL,
+		Adapter:       req.Adapter,
+		Models:        chat,
+		ImageModels:   image,
+		Endpoint:      res.Endpoint,
+		ImageEndpoint: res.ImageEndpoint,
+		ModelsPath:    res.ModelsPath,
+		Tried:         res.Tried,
+		LatencyMS:     res.LatencyMS,
+		Note:          res.Note,
+		Warnings:      warns,
+	}
+	if len(chat) == 0 && len(image) == 0 {
+		out.Warnings = append(out.Warnings,
+			"every discovered model was excluded; applying this would create a source with no models")
+	}
+
+	if !apply {
+		writeJSON(w, http.StatusOK, out)
+		return
+	}
+
+	name := strings.TrimSpace(req.Name)
+	if name == "" {
+		name = defaultSourceName(req.BaseURL)
+	}
+	if name == "" {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			"name required: cannot derive one from base_url")
+		return
+	}
+	if strings.ContainsAny(name, "/\\ \t") || name == "." || name == ".." {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			"source name must not contain slashes or whitespace")
+		return
+	}
+
+	exists := g.sourceExists(name)
+	if exists && req.ReplaceModels == nil {
+		// Refusing here is deliberate. Re-running discovery on an existing
+		// source and having it silently replace a hand-tuned model list —
+		// including priorities chosen for the AUTO chain — is the kind of
+		// "helpful" behaviour that loses an afternoon.
+		writeError(w, http.StatusConflict, "conflict",
+			"source "+name+" already exists; set replace_models=true to overwrite its model list, "+
+				"or choose a different name")
+		return
+	}
+
+	// Build the source directly rather than round-tripping through the
+	// /api/sources payload: resolveAPIKey carries the "an empty api_key means
+	// keep the stored one" rule, and going through a second decoder here would
+	// duplicate (and eventually diverge from) that logic.
+	src := config.Source{
+		Name:          name,
+		BaseURL:       res.BaseURL,
+		APIKey:        g.resolveAPIKey(name, req.APIKey, exists),
+		Adapter:       req.Adapter,
+		Endpoint:      res.Endpoint,
+		ImageEndpoint: res.ImageEndpoint,
+		Models:        modelsToConfig(chat, image),
+		Temperature:   req.Temperature,
+	}
+	if req.MaxConcurrent > 0 {
+		src.MaxConcurrent = req.MaxConcurrent
+	}
+
+	// An existing source may carry fields discovery does not know about
+	// (proxy_url, headers, api_key_env, timeout). Rebuilding from scratch would
+	// drop them, so they are inherited unless the request overrides them.
+	if exists {
+		old := g.sourceByName(name)
+		if old != nil {
+			src.ProxyURL = old.ProxyURL
+			src.APIKeyEnv = old.APIKeyEnv
+			src.Timeout = old.Timeout
+			src.QueueTimeout = old.QueueTimeout
+			src.Headers = old.Headers
+			src.Meta = old.Meta
+			src.MaxTokens = old.MaxTokens
+			src.RPM = old.RPM
+			// image_endpoint is NOT inherited when discovery supplied one:
+			// the stored value was correct for the OLD base_url, and the base
+			// is exactly what may have changed. Inheriting it would keep a
+			// stale value that 404s against the new base.
+			if res.ImageEndpoint == "" {
+				src.ImageEndpoint = old.ImageEndpoint
+			}
+			if src.Adapter == "" {
+				src.Adapter = old.Adapter
+			}
+			if req.MaxConcurrent <= 0 {
+				src.MaxConcurrent = old.MaxConcurrent
+			}
+			// replace_models=false keeps the curated list. Discovery still
+			// refreshed base_url/endpoint/credential, which is the point of
+			// re-running it.
+			if req.ReplaceModels != nil && !*req.ReplaceModels {
+				src.Models = old.Models
+				out.Warnings = append(out.Warnings,
+					"model list left unchanged (replace_models=false)")
+			}
+		}
+	}
+
+	if err := g.core.AddSource(src); err != nil {
+		writeError(w, http.StatusBadRequest, "source_error", err.Error())
+		return
+	}
+
+	out.SourceName = name
+	out.Applied = true
+	writeJSON(w, http.StatusOK, out)
+}
+
+// selectModels applies the include/exclude filters.
+//
+// An include entry the catalog does not have is an error: silently dropping it
+// produces a source that looks configured and 404s at request time, which is
+// far harder to diagnose than a rejected apply. An exclude entry that is not
+// present is fine — the operator is allowed to be defensive.
+func selectModels(chat, image, include, exclude []string) ([]string, []string, []string, error) {
+	have := map[string]bool{}
+	for _, m := range chat {
+		have[m] = true
+	}
+	for _, m := range image {
+		have[m] = true
+	}
+
+	if len(include) > 0 {
+		var missing []string
+		for _, want := range include {
+			if !have[want] {
+				missing = append(missing, want)
+			}
+		}
+		if len(missing) > 0 {
+			return nil, nil, nil, &discoverError{
+				msg: "these models are not in the upstream catalog: " + strings.Join(missing, ", "),
+			}
+		}
+		keep := map[string]bool{}
+		for _, want := range include {
+			keep[want] = true
+		}
+		var c, i []string
+		for _, m := range chat {
+			if keep[m] {
+				c = append(c, m)
+			}
+		}
+		for _, m := range image {
+			if keep[m] {
+				i = append(i, m)
+			}
+		}
+		chat, image = c, i
+	}
+
+	if len(exclude) > 0 {
+		drop := map[string]bool{}
+		for _, m := range exclude {
+			drop[m] = true
+		}
+		var c, i []string
+		for _, m := range chat {
+			if !drop[m] {
+				c = append(c, m)
+			}
+		}
+		for _, m := range image {
+			if !drop[m] {
+				i = append(i, m)
+			}
+		}
+		chat, image = c, i
+	}
+
+	// chat and image must stay non-nil: the JSON contract forbids null here.
+	if chat == nil {
+		chat = []string{}
+	}
+	if image == nil {
+		image = []string{}
+	}
+
+	var warns []string
+	if len(chat) == 0 && len(image) == 0 {
+		// Caller can only get here by excluding everything, or from a catalog
+		// that was already empty. Either way "silence" would read as success.
+		warns = append(warns,
+			"no models selected; applying this would create a source that cannot serve any request")
+		return chat, image, warns, nil
+	}
+	if len(image) > 0 && len(chat) == 0 {
+		warns = append(warns,
+			"every model is an image model; AUTO chat requests have nothing to select")
+	}
+	if len(chat) > 0 && len(image) > 0 {
+		warns = append(warns,
+			"image models are only reachable through /v1/images/generations, not AUTO chat")
+	}
+	return chat, image, warns, nil
+}
+
+// discoverError is a client mistake rather than an upstream failure, which is
+// why it gets its own type instead of fmt.Errorf.
+type discoverError struct{ msg string }
+
+func (e *discoverError) Error() string { return e.msg }
+
+// defaultSourceName derives a source name from a base URL host, stripping the
+// version segment and any "api." prefix. It returns "" for input that is not a
+// URL at all — the caller must not fall back to echoing garbage into a source
+// name, since the name becomes a config key and an API path segment.
+func defaultSourceName(baseURL string) string {
+	s := strings.TrimSpace(baseURL)
+	if !strings.Contains(s, "://") {
+		return ""
+	}
+	s = s[strings.Index(s, "://")+3:]
+	if i := strings.IndexAny(s, "/?#"); i >= 0 {
+		s = s[:i]
+	}
+	if i := strings.LastIndex(s, ":"); i > 0 && !strings.Contains(s[i+1:], "]") {
+		s = s[:i] // strip port
+	}
+	// A host must still look like one after trimming the cosmetic prefixes.
+	s = strings.TrimPrefix(s, "api.")
+	s = strings.TrimPrefix(s, "www.")
+	if s == "" || strings.ContainsAny(s, " \t\"'<>") {
+		return ""
+	}
+	for _, r := range s {
+		switch {
+		case r >= 'a' && r <= 'z', r >= 'A' && r <= 'Z', r >= '0' && r <= '9':
+		case r == '.' || r == '-' || r == '_':
+		default:
+			return ""
+		}
+	}
+	if !strings.Contains(s, ".") {
+		// A bare label is far more likely to be a mistake than a hostname.
+		return ""
+	}
+	return s
+}

internal/gateway/discover_api_test.go

@@ -0,0 +1,314 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"modelrouter/internal/config"
+)
+
+// upstream that answers /v1/models with the real captured catalog
+func catalogUpstream(t *testing.T) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		if got := r.Header.Get("Authorization"); got != "Bearer sk-secret" {
+			w.WriteHeader(http.StatusUnauthorized)
+			return
+		}
+		_, _ = w.Write([]byte(realCatalogBody))
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestDefaultSourceName(t *testing.T) {
+	cases := map[string]string{
+		"https://api.vendor.com/v1":          "vendor.com",
+		"https://api.openai.com/v1":          "openai.com",
+		"https://www.vendor.com/v1":          "vendor.com",
+		"http://127.0.0.1:8080/v1":           "127.0.0.1",
+		"https://amr-link.open-design.ai/v1": "amr-link.open-design.ai",
+		"not a url at all":                   "",
+	}
+	for in, want := range cases {
+		if got := defaultSourceName(in); got != want {
+			t.Errorf("defaultSourceName(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestSelectModels_IncludeUnknownIsAnError(t *testing.T) {
+	chat := []string{"a", "b"}
+	image := []string{"img"}
+	// Silently dropping a typo'd id would create a source that looks
+	// configured and 404s on every request — much worse than a rejection.
+	_, _, _, err := selectModels(chat, image, []string{"a", "typo"}, nil)
+	if err == nil {
+		t.Fatal("expected an error for an include id that is not in the catalog")
+	}
+	if !strings.Contains(err.Error(), "typo") {
+		t.Errorf("error should name the offending id, got %q", err)
+	}
+}
+
+func TestSelectModels_IncludeNarrows(t *testing.T) {
+	chat, image, _, err := selectModels([]string{"a", "b", "c"}, []string{"img1", "img2"}, []string{"b", "img1"}, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(chat) != 1 || chat[0] != "b" {
+		t.Errorf("chat = %v, want [b]", chat)
+	}
+	if len(image) != 1 || image[0] != "img1" {
+		t.Errorf("image = %v, want [img1]", image)
+	}
+}
+
+func TestSelectModels_ExcludeUnknownIsFine(t *testing.T) {
+	chat, _, _, err := selectModels([]string{"a"}, nil, nil, []string{"never-existed"})
+	if err != nil {
+		t.Errorf("excluding a model that is not present should be allowed: %v", err)
+	}
+	if len(chat) != 1 {
+		t.Errorf("chat = %v, want [a] unchanged", chat)
+	}
+}
+
+func TestSelectModels_NeverReturnsNil(t *testing.T) {
+	chat, image, warns, err := selectModels(nil, nil, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if chat == nil || image == nil {
+		t.Fatal("nil slices would serialize as JSON null and break the UI")
+	}
+	if len(warns) == 0 {
+		t.Error("an empty catalog deserves a warning, not silence")
+	}
+}
+
+func TestSelectModels_WarnsAboutImageOnlyCatalog(t *testing.T) {
+	_, _, warns, err := selectModels(nil, []string{"gpt-image-2"}, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	joined := strings.Join(warns, " ")
+	if !strings.Contains(joined, "AUTO") {
+		t.Errorf("warnings %v should say AUTO has nothing to select", warns)
+	}
+}
+
+func TestSelectModels_WarnsThatImagesAreNotInAUTO(t *testing.T) {
+	_, _, warns, err := selectModels([]string{"a"}, []string{"img"}, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(strings.Join(warns, " "), "images/generations") {
+		t.Errorf("warnings %v should point at the image endpoint", warns)
+	}
+}
+
+// TestDiscoveryResultAndConfigRoundTrip checks the fields the WebUI will echo
+// back actually survive into a config.Source — a mismatch here means the
+// operator previews one set of values and saves another.
+func TestDiscoveryResultAndConfigRoundTrip(t *testing.T) {
+	res := &discoveryResult{
+		BaseURL:     "https://h",
+		Endpoint:    "/v1/chat/completions",
+		Models:      []string{"m1", "m2"},
+		ImageModels: []string{"img1"},
+	}
+	src := config.Source{
+		Name:     "x",
+		BaseURL:  res.BaseURL,
+		Endpoint: res.Endpoint,
+		Models:   modelsToConfig(res.Models, res.ImageModels),
+	}
+	if len(src.Models) != 3 {
+		t.Fatalf("models = %d, want 3", len(src.Models))
+	}
+	if src.Endpoint != "/v1/chat/completions" {
+		t.Errorf("endpoint lost: %q", src.Endpoint)
+	}
+	var img int
+	for _, m := range src.Models {
+		if m.Kind == "image" {
+			img++
+		}
+	}
+	if img != 1 {
+		t.Errorf("image models = %d, want 1", img)
+	}
+}
+
+// TestDiscoverModels_SendsAuthorizationHeader pins the only way the feature
+// works against a real vendor: without the bearer token every paid /models
+// endpoint returns 401.
+func TestDiscoverModels_SendsAuthorizationHeader(t *testing.T) {
+	var seen string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		seen = r.Header.Get("Authorization")
+		_, _ = w.Write([]byte(`{"data":[{"id":"m"}]}`))
+	}))
+	defer srv.Close()
+
+	if _, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-secret"); err != nil {
+		t.Fatal(err)
+	}
+	if seen != "Bearer sk-secret" {
+		t.Errorf("Authorization = %q, want %q", seen, "Bearer sk-secret")
+	}
+}
+
+// TestDiscoverModels_WrongKeyIsUnauthorizedNotBadGateway pins the distinction
+// that decides whether the operator knows what to do next.
+//
+// A rejected key and an unreachable vendor are the two most common discovery
+// failures, and they need opposite actions: re-type the key, versus go check the
+// vendor's status page. Both used to surface as one generic error, so the
+// operator had to guess.
+func TestDiscoverModels_WrongKeyIsUnauthorizedNotBadGateway(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusUnauthorized)
+		_, _ = w.Write([]byte(`{"error":{"message":"invalid api key","code":"invalid_api_key"}}`))
+	}))
+	defer srv.Close()
+
+	_, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-wrong")
+	if err == nil {
+		t.Fatal("expected a failure against a 401 endpoint")
+	}
+	var f *discoveryFailure
+	if !errors.As(err, &f) {
+		t.Fatalf("error should be classified, got %T: %v", err, err)
+	}
+	if got := f.StatusOrDefault(); got != http.StatusUnauthorized {
+		t.Errorf("status = %d, want %d — a wrong key must not read as an upstream outage",
+			got, http.StatusUnauthorized)
+	}
+	// The message has to name which credential is at fault, or the status code
+	// alone still sends the operator to the wrong place.
+	if !strings.Contains(strings.ToLower(f.Error()), "api key") {
+		t.Errorf("message %q should name the api key as the cause", f.Error())
+	}
+	if !strings.Contains(f.Error(), "/v1/models") {
+		t.Errorf("message %q should list what was tried", f.Error())
+	}
+}
+
+// TestDiscoverModels_UnreachableStaysBadGateway: the classification must not
+// over-correct — an unresolvable host is still the vendor's or the network's
+// problem, not the key's.
+func TestDiscoverModels_UnreachableStaysBadGateway(t *testing.T) {
+	_, err := DiscoverModels(context.Background(), "https://no-such-host-xyz.invalid/v1", "k")
+	if err == nil {
+		t.Fatal("expected a failure for an unresolvable host")
+	}
+	var f *discoveryFailure
+	if !errors.As(err, &f) {
+		t.Fatalf("error should be classified, got %T", err)
+	}
+	if got := f.StatusOrDefault(); got != http.StatusBadGateway {
+		t.Errorf("status = %d, want %d", got, http.StatusBadGateway)
+	}
+	if !strings.Contains(strings.ToLower(f.Error()), "resolve") {
+		t.Errorf("message %q should say the hostname does not resolve", f.Error())
+	}
+}
+
+// TestDiscoveryCandidates_NoDuplicateRequests: base "https://h/v1" with path
+// "/models" and base "https://h" with path "/v1/models" request the SAME URL.
+// Listing it twice in a failure message reads as "two things were tried" when
+// only one was.
+func TestDiscoveryCandidates_NoDuplicateRequests(t *testing.T) {
+	seen := map[string]bool{}
+	for _, c := range discoveryCandidates("https://h/v1") {
+		url := c.baseURL + c.modelsPath
+		if seen[url] {
+			t.Errorf("candidate %q would be requested twice", url)
+		}
+		seen[url] = true
+	}
+	if len(seen) < 2 {
+		t.Errorf("only %d distinct URLs; the search collapsed", len(seen))
+	}
+}
+
+// TestDiscoverModels_FailureListsEachURLOnce is the end-to-end form of the same
+// rule, on the error text an operator actually reads.
+func TestDiscoverModels_FailureListsEachURLOnce(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv.Close()
+
+	_, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err == nil {
+		t.Fatal("expected failure")
+	}
+	msg := err.Error()
+	needle := srv.URL + "/v1/models"
+	if n := strings.Count(msg, needle); n > 1 {
+		t.Errorf("URL %s appears %d times in %q", needle, n, msg)
+	}
+}
+
+func TestDiscoverModels_EmptyKeyOmitsHeader(t *testing.T) {
+	var present bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, present = r.Header["Authorization"]
+		_, _ = w.Write([]byte(`{"data":[{"id":"m"}]}`))
+	}))
+	defer srv.Close()
+
+	if _, err := DiscoverModels(context.Background(), srv.URL+"/v1", ""); err != nil {
+		t.Fatal(err)
+	}
+	if present {
+		t.Error("an empty key should omit the header, not send an empty bearer")
+	}
+}
+
+// TestDiscoverModels_RealCatalogEndToEnd walks the actual captured response
+// through the whole path and asserts the numbers an operator would see.
+func TestDiscoverModels_RealCatalogEndToEnd(t *testing.T) {
+	srv := catalogUpstream(t)
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-secret")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL+"/v1" {
+		t.Errorf("base = %q, want the typed URL unchanged", res.BaseURL)
+	}
+	if res.Endpoint != "" {
+		t.Errorf("endpoint = %q, want empty (base already carries /v1)", res.Endpoint)
+	}
+	if len(res.Models) != 2 || len(res.ImageModels) != 2 {
+		t.Errorf("chat/image = %d/%d, want 2/2", len(res.Models), len(res.ImageModels))
+	}
+	if res.Adapter != "openai" {
+		t.Errorf("adapter = %q, want openai", res.Adapter)
+	}
+
+	// The response must be JSON the UI can consume directly.
+	b, _ := json.Marshal(res)
+	var round map[string]interface{}
+	if err := json.Unmarshal(b, &round); err != nil {
+		t.Fatal(err)
+	}
+	if round["ok"] != nil {
+		// discoverResponse adds ok; discoveryResult does not. Both marshal
+		// fine, so this is only a shape note.
+		_ = round
+	}
+}

internal/gateway/discover_endpoint_test.go

@@ -0,0 +1,153 @@
+package gateway
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// TestComposeImageEndpoint_NoVersionBaseLeavesDefault is the counterpart of the
+// /v1 case: when the base carries no version segment, discovery must emit "" so
+// the gateway's own default "/v1/images/generations" composes.
+//
+// This is the shape every currently-working source uses — siliconflow, the only
+// source in production with kind:image, has base_url
+// "https://api.siliconflow.cn" — so writing a non-empty value here would break
+// exactly the configuration that is known to work today.
+func TestComposeImageEndpoint_NoVersionBaseLeavesDefault(t *testing.T) {
+	if got := composeImageEndpoint("https://api.siliconflow.cn"); got != "" {
+		t.Errorf("composeImageEndpoint = %q, want \"\" so the built-in default composes", got)
+	}
+	// And the composition it implies must be the vendor's real path.
+	base := "https://api.siliconflow.cn"
+	composed := base + "/v1/images/generations"
+	if !strings.HasSuffix(composed, "/v1/images/generations") {
+		t.Errorf("composed URL %q lost the image path", composed)
+	}
+}
+
+// TestDiscovery_BareHostNeedsVersionOnEndpointWhenCatalogIsVersioned
+// distinguishes two cases that look identical from the outside.
+//
+//	bare host + catalog at "/models"       → chat is "/chat/completions" (the
+//	  adapter default, no endpoint written)
+//	bare host + catalog at "/v1/models"    → chat is "/v1/chat/completions",
+//	  so an endpoint MUST be written
+//
+// The catalog path is the only evidence of whether the upstream wants the
+// version segment, and getting this backwards sends every chat request to a
+// path the vendor does not serve.
+func TestDiscovery_BareHostNeedsVersionOnEndpointWhenCatalogIsVersioned(t *testing.T) {
+	cases := []struct {
+		name         string
+		catalogPath  string
+		wantEndpoint string
+		wantImage    string
+	}{
+		{
+			name:         "catalog at the bare /models needs no endpoint",
+			catalogPath:  "/models",
+			wantEndpoint: "",
+			wantImage:    "",
+		},
+		{
+			name:         "catalog at /v1/models requires /v1 on the chat endpoint",
+			catalogPath:  "/v1/models",
+			wantEndpoint: "/v1/chat/completions",
+			wantImage:    "",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if r.URL.Path != tc.catalogPath {
+					w.WriteHeader(http.StatusNotFound)
+					return
+				}
+				_, _ = w.Write([]byte(`{"data":[{"id":"m1"},{"id":"m2"}]}`))
+			}))
+			defer srv.Close()
+
+			res, err := DiscoverModels(context.Background(), srv.URL, "k")
+			if err != nil {
+				t.Fatalf("discovery failed: %v", err)
+			}
+			if res.BaseURL != srv.URL {
+				t.Errorf("base = %q, want %q", res.BaseURL, srv.URL)
+			}
+			if res.Endpoint != tc.wantEndpoint {
+				t.Errorf("endpoint = %q, want %q", res.Endpoint, tc.wantEndpoint)
+			}
+			if res.ImageEndpoint != tc.wantImage {
+				t.Errorf("image_endpoint = %q, want %q", res.ImageEndpoint, tc.wantImage)
+			}
+			// Whatever it wrote, the chat URL must reach the same version
+			// prefix the catalog was served under.
+			ep := res.Endpoint
+			if ep == "" {
+				ep = "/chat/completions"
+			}
+			if got := res.BaseURL + ep; strings.Contains(got, "/v1/v1") {
+				t.Errorf("chat URL %q has a doubled version", got)
+			}
+		})
+	}
+}
+
+// TestComposeEndpoints_BothFormsProduceWorkingURLs is the end-to-end
+// invariant across the two shapes, stated as the URLs the gateway will build.
+//
+// A base with a version and no image_endpoint produces "/v1/v1/…" — the exact
+// 404 observed before composeImageEndpoint existed. A base without one and a
+// non-empty image_endpoint produces a path missing the version. Both are
+// silent, expensive mistakes, so they are asserted literally.
+func TestComposeEndpoints_BothFormsProduceWorkingURLs(t *testing.T) {
+	cases := []struct {
+		name      string
+		base      string
+		wantChat  string
+		wantImage string
+	}{
+		{
+			name:     "bare host: adapter and image defaults compose",
+			base:     "https://h",
+			wantChat: "https://h/chat/completions",
+			// openai's adapter endpoint is /chat/completions (verified in
+			// internal/lua/adapters/openai.lua: adapter.endpoint).
+			wantImage: "https://h/v1/images/generations",
+		},
+		{
+			name:      "versioned base: the version moves onto both endpoints",
+			base:      "https://h/v1",
+			wantChat:  "https://h/v1/chat/completions",
+			wantImage: "https://h/v1/images/generations",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			ep := composeEndpoint("/models", tc.base, tc.base)
+			if ep == "" {
+				ep = "/chat/completions" // the adapter default
+			}
+			iep := composeImageEndpoint(tc.base)
+			if iep == "" {
+				iep = "/v1/images/generations" // the gateway default
+			}
+			gotChat := tc.base + ep
+			gotImage := tc.base + iep
+			if gotChat != tc.wantChat {
+				t.Errorf("chat URL = %q, want %q", gotChat, tc.wantChat)
+			}
+			if gotImage != tc.wantImage {
+				t.Errorf("image URL = %q, want %q", gotImage, tc.wantImage)
+			}
+			for _, u := range []string{gotChat, gotImage} {
+				if strings.Contains(u, "/v1/v1") {
+					t.Errorf("%q has a doubled version segment", u)
+				}
+			}
+		})
+	}
+}

internal/gateway/discover_mutate_test.go

@@ -0,0 +1,347 @@
+package gateway
+
+// Real mutation injection: rewrites the IMPLEMENTATION on disk, re-runs the
+// suite, and requires it to go red.
+//
+// Why this exists and why it is a Go file rather than a shell script:
+//
+// The first version was a shell script. Its -run filter contained a regex that
+// the shell mangled, so every mutant reported "SURVIVED" while go test printed
+// "[no tests to run]". Four mutants looked like judgement failures and were
+// actually never executed. A judgement that cannot tell "tests passed" from
+// "no tests ran" is not a judgement.
+//
+// This version therefore refuses to report anything unless it has first proved
+// the filter selects a non-trivial number of tests (see baseline below), and it
+// treats "[no tests to run]" as a hard error rather than a green run.
+//
+// It also must not be run against a DIRTY tree: the mutant is written over
+// discover.go, and a leftover mutant from a killed run would make the baseline
+// green against broken code while every verdict came from the wrong source.
+// See the guard at the top of TestRealMutants.
+//
+// Run: go test -tags luajit ./internal/gateway/ -run TestRealMutants -v
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sort"
+	"strings"
+	"testing"
+)
+
+// mutationFilter must select every assertion that guards the code below. It is
+// deliberately broad: a narrow filter is how mutants get to survive for free.
+const mutationFilter = "Test(ParseModelCatalog|DiscoveredModelKind|Discover|TrimVersion|VersionSegment|ComposeEndpoint|ComposeImageEndpoint|ComposeEndpoints|IsVersionSegment|ModelsToConfig|DiscoveryResult|DiscoveryCandidates|Discovery_BareHost)"
+
+type mutation struct {
+	name string
+	from string
+	to   string
+}
+
+func TestRealMutants(t *testing.T) {
+	if testing.Short() {
+		t.Skip("mutation campaign compiles the package repeatedly")
+	}
+	wd, err := os.Getwd()
+	if err != nil {
+		t.Fatal(err)
+	}
+	target := filepath.Join(wd, "discover.go")
+	orig, err := os.ReadFile(target)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Refuse to run against a dirty tree.
+	//
+	// The mutant is written OVER discover.go, so on a dirty file the "original"
+	// captured above is not the committed source — the baseline then runs against
+	// whatever was there, and a killed run (a background-task timeout, a Ctrl-C)
+	// leaves a mutant behind that the NEXT run treats as the baseline. Both make
+	// every verdict meaningless while the harness still reports a clean score.
+	//
+	// Refusing rather than restoring: this cannot tell "leftover mutant" apart from
+	// "the developer is mid-edit", and silently reverting real work is worse than
+	// stopping.
+	if dirty, derr := mutationTargetIsDirty(wd, "discover.go"); derr == nil && dirty {
+		t.Fatalf("%s has uncommitted changes — commit or stash them first. A mutant "+
+			"left behind by a killed run would otherwise become the baseline, and "+
+			"every verdict below would be about code nobody wrote.", target)
+	}
+	defer func() {
+		if werr := os.WriteFile(target, orig, 0o644); werr != nil {
+			t.Fatalf("restore discover.go: %v", werr)
+		}
+	}()
+
+	// --- prove the harness is armed before trusting any verdict ---
+	// A filter that selects nothing exits 0, which is indistinguishable from a
+	// real pass. The only way to tell them apart is to compare against a filter
+	// KNOWN to be empty: if that empty filter produces the same result as the
+	// real one, the real one is probably empty too.
+	//
+	// Note go test does NOT print "[no tests to run]" for a wholly unmatched
+	// filter (it prints that only when the package has other matching tests), so
+	// the check has to compare outputs rather than look for a marker string.
+	// Getting this backwards made the harness refuse to run at all.
+	emptyOut := runSuite(wd, "TestZZZNoSuchTestExists")
+	realOut := runSuite(wd, mutationFilter)
+	if countTests(emptyOut) != 0 {
+		t.Fatalf("a filter matching nothing selected %d tests; the harness's own probe is broken",
+			countTests(emptyOut))
+	}
+	// Kept for the no-op check below: a mutant whose output matches the
+	// baseline never changed behaviour, so it is an invalid mutation rather
+	// than a gap in the tests.
+	baseOut := realOut
+	if strings.Contains(realOut, "[no tests to run]") {
+		t.Fatalf("baseline selected no tests:\n%s", realOut)
+	}
+	baseCount := countTests(realOut)
+	if baseCount < 25 {
+		t.Fatalf("baseline ran only %d tests; the filter is too narrow to judge anything", baseCount)
+	}
+	if strings.Contains(realOut, "FAIL") {
+		t.Logf("note: baseline is not green; kill verdicts remain valid")
+	}
+	t.Logf("baseline: %d tests selected, empty filter selects %d",
+		baseCount, countTests(emptyOut))
+
+	muts := []mutation{
+		{
+			// Written as a condition that still READS modOut, because a
+			// mutation that leaves a variable unused fails to compile — and a
+			// build failure is not a judgement (see the harness note above).
+			name: "image kind ignores upstream output modality",
+			from: `if m.Capabilities.ImageGeneration || has(modOut, "image") {`,
+			to:   "if m.Capabilities.ImageGeneration || (len(modOut) > 0 && false) {",
+		},
+		{
+			name: "kind defaults to image (guessing from names)",
+			from: "\t}\n\treturn \"chat\"\n}\n\n// modalityList",
+			to:   "\t}\n\treturn \"image\"\n}\n\n// modalityList",
+		},
+		{
+			name: "nested metadata no longer read",
+			// 锚点必须与 discover.go 逐字节一致：此前这里是 6 个空格，源码已
+			// 变成 2 个空格，于是这条变异永远"找不到锚点"——变异campaign 每次
+			// 都以 anchor not found 失败，而它本该验证"嵌套 metadata 被丢掉"能被
+			// 测试抓住。锚点漂移 = 这条变异从未真正跑过（和 issue #1 里
+			// `cc["models"]` 恒 false 的测试跳过是同一种失效：断言写了但没执行）。
+			from: "\tcopyFields(m.Metadata)  // nested first",
+			to:   "\tcopyFields(json.RawMessage(nil)) // mutant: nested metadata dropped",
+		},
+		{
+			name: "non-catalog bodies accepted",
+			from: "\tif trimmed == \"\" || trimmed == \"null\" {\n\t\treturn nil, fmt.Errorf(\"empty response\")\n\t}",
+			to:   "\tif false && (trimmed == \"\" || trimmed == \"null\") {\n\t\treturn nil, fmt.Errorf(\"empty response\")\n\t}",
+		},
+		{
+			// Removing the bare "/models" candidate must be observable: a vendor
+			// serving its catalog at the bare path would otherwise look
+			// unreachable. The `len(...) >= 0` is a tautology that only exists to
+			// keep the expression syntactically valid after the call is dropped.
+			name: "URL fallback search removed",
+			from: "\t\tadd(raw, \"/models\", 0)",
+			to:   "\t\tif len(\"/models\") >= 0 {\n\t\t\t_ = hasVersion\n\t\t}",
+		},
+		{
+			name: "version segment never moved onto the chat endpoint",
+			from: "\tif version == \"\" {\n\t\treturn \"\" // adapter default already composes\n\t}",
+			to:   "\tif version == \"\" || len(version) > 0 {\n\t\treturn \"\"\n\t}",
+		},
+		{
+			name: "image endpoint never corrected (/v1/v1 regression)",
+			from: "\tif b := trimVersionSuffix(persistedBase); b != persistedBase {\n\t\treturn \"/images/generations\"\n\t}",
+			to:   "\tif b := trimVersionSuffix(persistedBase); b == b && b != b {\n\t\treturn \"/images/generations\"\n\t}",
+		},
+		{
+			name: "version test loosened so 'vision' counts as a version",
+			from: "\tsuffix := strings.ToLower(seg[i:])",
+			to:   "\tsuffix := strings.ToLower(seg[i:])\n\tif len(suffix) >= 0 {\n\t\tsuffix = \"beta\"\n\t}",
+		},
+		{
+			// The guard must be removable WITHOUT the mutant becoming a
+			// semantic no-op. Two earlier attempts failed:
+			//
+			//   `&& out.BaseURL != ""`  — the guard this test exercises uses a
+			//     zero-value struct whose BaseURL is empty, so the added
+			//     condition made the mutant behave identically to the original.
+			//   `&& len(out.BaseURL) >= 0` — always true for any string, so the
+			//     mutant was STILL an identity transform.
+			//
+			// Reading the variable into a blank assignment keeps the compiler
+			// happy without changing behaviour, which is what makes the
+			// mutant observable.
+			name: "nil slice guard removed from JSON output",
+			from: "\tif out.ImageModels == nil {\n\t\tout.ImageModels = []string{}\n\t}",
+			to:   "\tif out.ImageModels == nil {\n\t\t_ = out.ImageModels\n\t}",
+		},
+		{
+			name: "auth rejection no longer classified",
+			from: "\tif status != http.StatusUnauthorized && status != http.StatusForbidden {\n\t\treturn \"\"\n\t}",
+			to:   "\tif status != http.StatusUnauthorized && status != http.StatusForbidden || len(body) >= 0 {\n\t\treturn \"\"\n\t}",
+		},
+		{
+			name: "duplicate candidate URLs no longer collapsed",
+			from: "\t\tif seenPair[pair] || seenURL[base+path] {",
+			to:   "\t\tif len(pair) > 0 && (seenPair[pair] || seenURL[base+path]) && false {",
+		},
+	}
+
+	survived := 0
+	for _, m := range muts {
+		t.Run(m.name, func(t *testing.T) {
+			if !strings.Contains(string(orig), m.from) {
+				t.Fatalf("anchor not found in discover.go — this mutation would be a no-op:\n%q", m.from)
+			}
+			mutated := strings.Replace(string(orig), m.from, m.to, 1)
+			if mutated == string(orig) {
+				t.Fatal("replacement changed nothing")
+			}
+			if werr := os.WriteFile(target, []byte(mutated), 0o644); werr != nil {
+				t.Fatal(werr)
+			}
+			defer os.WriteFile(target, orig, 0o644)
+
+			out := runSuite(wd, mutationFilter)
+			// A build failure kills EVERY mutant, so it would mask a broken
+			// harness as a perfect score. Distinguish it from a real test
+			// failure and fail loudly instead.
+			if strings.Contains(out, "[build failed]") {
+				t.Fatalf("mutant broke the build — that is not a judgement, it is a broken harness:\n%s", firstLines(out, 20))
+			}
+			if countTests(out) == 0 {
+				t.Fatalf("filter selected nothing — the verdict would be meaningless")
+			}
+			if !strings.Contains(out, "FAIL") {
+				// The mutant compiled and the suite stayed green. Before calling
+				// this a survivor, check the mutation was not a no-op: an earlier
+				// version reported four "survivors" that were really invalid
+				// mutations whose added guard made them behaviourally identical
+				// to the original. A silent no-op is a broken mutation, and
+				// conflating the two sends the next reader hunting for a missing
+				// test that does not exist.
+				if behaviourOnly(out) == behaviourOnly(baseOut) {
+					t.Fatalf("mutant is a no-op: the suite output is identical to the "+
+						"baseline, so this mutation never changed behaviour. Fix the "+
+						"mutation rather than looking for a missing test.\nmutant: %q", m.to)
+				}
+				survived++
+				t.Errorf("SURVIVED %s — no test noticed this wrong implementation", m.name)
+				return
+			}
+			t.Logf("KILL  %s", m.name)
+		})
+	}
+	if survived > 0 {
+		t.Errorf("%d of %d mutants survived", survived, len(muts))
+	}
+}
+
+// runSuite invokes the go tool on this package with the given -run filter.
+//
+// The luajit build tag is mandatory: without it the package falls back to plain
+// Lua 5.1, billing.lua's `goto` fails to compile, and the whole package fails
+// to build — which the harness would misread as "every mutant killed".
+func runSuite(wd, filter string) string {
+	tool := os.Getenv("GO_TEST_GO")
+	if tool == "" {
+		tool = "go"
+	}
+	cmd := exec.Command(tool, "test",
+		"-tags", "luajit",
+		".", "-run", filter, "-count=1", "-v")
+	cmd.Dir = wd
+	cmd.Env = os.Environ()
+	out, _ := cmd.CombinedOutput()
+	return string(out)
+}
+
+// mutationTargetIsDirty reports whether a file under internal/gateway has
+// uncommitted changes, so the harness can refuse to run instead of treating a
+// leftover mutant as the baseline.
+//
+// git runs from the package directory with a PATH RELATIVE to it. An earlier
+// version of the sibling script passed a repo-relative pathspec while running in
+// the package directory; git matched nothing and the guard never fired.
+func mutationTargetIsDirty(wd, name string) (bool, error) {
+	cmd := exec.Command("git", "diff", "--name-only", "--", "./"+name)
+	cmd.Dir = wd
+	out, err := cmd.Output()
+	if err != nil {
+		return false, err
+	}
+	return strings.TrimSpace(string(out)) != "", nil
+}
+
+func countTests(out string) int {
+	n := 0
+	for _, line := range strings.Split(out, "\n") {
+		// Subtests are indented; only count top-level ones so a single test
+		// with a table of cases cannot inflate the count.
+		if strings.HasPrefix(line, "=== RUN   Test") && !strings.Contains(line, "/") {
+			n++
+		}
+	}
+	return n
+}
+
+// behaviourOnly strips the parts of `go test -v` output that vary between runs
+// (timings, goroutine ids, temp paths) so two runs can be compared for equality.
+// If a mutant produces output identical to the baseline under this projection,
+// the mutation did not change observable behaviour and is therefore invalid.
+func behaviourOnly(out string) string {
+	var kept []string
+	for _, line := range strings.Split(out, "\n") {
+		trimmed := strings.TrimSpace(line)
+		switch {
+		// Keep only the outcome lines. Test NAMES are deliberately dropped:
+		// two runs of the same suite can legitimately order or cache tests
+		// differently, and keeping the names made every mutant look like it had
+		// "changed behaviour" — which defeats the no-op detector entirely.
+		case strings.HasPrefix(trimmed, "--- PASS"), strings.HasPrefix(trimmed, "--- FAIL"),
+			strings.HasPrefix(trimmed, "--- SKIP"),
+			strings.HasPrefix(trimmed, "ok  "), trimmed == "FAIL",
+			strings.Contains(trimmed, ".go:"):
+			kept = append(kept, trimmed)
+		}
+	}
+	// Counting outcomes is stronger than comparing the sequence: a mutation
+	// that swaps which test fails still changes the multiset, and one that
+	// changes nothing cannot.
+	sort.Strings(kept)
+	return strings.Join(kept, "\n")
+}
+
+// firstLines trims output for an error message; a failing nested run can be
+// thousands of lines and the cause is always near the top.
+func firstLines(s string, n int) string {
+	lines := strings.Split(s, "\n")
+	if len(lines) > n {
+		lines = lines[:n]
+	}
+	return strings.Join(lines, "\n")
+}
+
+// TestMutationFilterCoversEveryGuard keeps the filter honest. Dropping a group
+// from it silently disarms that area's mutants.
+func TestMutationFilterCoversEveryGuard(t *testing.T) {
+	for _, group := range []string{
+		"ParseModelCatalog",   // nested metadata + non-catalog rejection
+		"DiscoveredModelKind", // kind follows upstream metadata
+		"Discover",            // URL fallback, auth classification
+		"ComposeEndpoint",     // version segment exactly once (chat)
+		"ComposeImageEndpoint",
+		"ComposeEndpoints", // the composed URLs, both base shapes
+		"IsVersionSegment", // "v1beta" vs "vision"
+		"DiscoveryResult",  // nil slices are arrays
+	} {
+		if !strings.Contains(mutationFilter, group) {
+			t.Errorf("mutation filter omits %s; mutants there survive silently", group)
+		}
+	}
+}

internal/gateway/discover_mutation_test.go

@@ -0,0 +1,171 @@
+package gateway
+
+// Mutation harness for the discovery code. Kept in the tree (not in a shell
+// script) so the -run filter is a Go string literal rather than something the
+// shell can mangle into "[no tests to run]" — which is exactly how an earlier
+// version of this harness reported four surviving mutants that were simply
+// never executed.
+//
+// Each mutant flips one behaviour; TestMutationHarness_Fails proves the
+// harness itself is armed.
+
+import (
+	"encoding/json"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+// discoveryTestFilter must cover every assertion that guards the behaviour
+// below. Narrowing it is how mutants get to "survive" for free.
+const discoveryTestFilter = "Test(ParseModelCatalog|DiscoveredModelKind|Discover|TrimVersion|VersionSegment|ComposeEndpoint|IsVersionSegment|ModelsToConfig|DiscoveryResult|DiscoveryCandidates)"
+
+func TestDiscoveryFilterRunsRealTests(t *testing.T) {
+	if discoveryTestFilter == "" {
+		t.Fatal("filter is empty; every mutant would survive trivially")
+	}
+	// The whole point: the catalog parser test must be inside the filter, or a
+	// mutant that stops reading nested metadata survives with a green run.
+	for _, group := range []string{
+		"ParseModelCatalog",   // nested metadata + non-catalog rejection
+		"DiscoveredModelKind", // kind must follow upstream metadata
+		"Discover",            // URL fallback + endpoint composition
+		"ComposeEndpoint",     // version segment exactly once
+		"IsVersionSegment",    // "v1beta" vs "vision"
+	} {
+		if !strings.Contains(discoveryTestFilter, group) {
+			t.Errorf("filter %q skips %s tests; mutants in that area survive", discoveryTestFilter, group)
+		}
+	}
+}
+
+// TestMutantsAreDetectable is the harness's own regression guard: it runs the
+// filter's worth of assertions against deliberately broken inputs and requires
+// each to be rejected. If a future edit makes one of these pass, the
+// corresponding protection is gone.
+func TestMutantsAreDetectable(t *testing.T) {
+	t.Run("image classification follows upstream metadata, not the model name", func(t *testing.T) {
+		// The mutant: drop the output-modality check, so every model is chat.
+		m := upstreamModel{ID: "gpt-image-2"}
+		m.Architecture.OutputModalities = []string{"image"}
+		if got := discoveredModelKind(m); got != "image" {
+			t.Errorf("mutant survived: got %q, want \"image\"", got)
+		}
+	})
+
+	t.Run("nested metadata is read", func(t *testing.T) {
+		var ms []upstreamModel
+		if err := json.Unmarshal([]byte(realCatalogBody), &ms); err == nil {
+			t.Fatal("fixture must be a wrapper object, not a bare array")
+		}
+		var wrapper struct {
+			Data []upstreamModel `json:"data"`
+		}
+		if err := json.Unmarshal([]byte(realCatalogBody), &wrapper); err != nil {
+			t.Fatal(err)
+		}
+		if len(wrapper.Data) == 0 {
+			t.Fatal("fixture parsed to zero models")
+		}
+		// The mutant: skip copyFields(m.Metadata) entirely.
+		meta := modelMeta(wrapper.Data[0])
+		if meta == nil {
+			t.Fatal("mutant survived: metadata is nil, so cost/context_limit are lost")
+		}
+		for _, k := range []string{"context_limit", "cost", "context_budget"} {
+			if _, ok := meta[k]; !ok {
+				t.Errorf("mutant survived: %s dropped from nested metadata", k)
+			}
+		}
+	})
+
+	t.Run("non-catalog responses are rejected", func(t *testing.T) {
+		for _, body := range []string{`<html>login</html>`, `null`, `{"error":{"message":"x"}}`, `{}`} {
+			if _, err := parseModelCatalog([]byte(body)); err == nil {
+				t.Errorf("mutant survived: %q accepted as a catalog", body)
+			}
+		}
+	})
+
+	t.Run("version segments are distinguished from real path segments", func(t *testing.T) {
+		if !isVersionSegment("v1beta") {
+			t.Error("mutant survived: v1beta not treated as a version")
+		}
+		for _, seg := range []string{"vision", "vertex", "v1x"} {
+			if isVersionSegment(seg) {
+				t.Errorf("mutant survived: %q wrongly treated as a version", seg)
+			}
+		}
+	})
+
+	t.Run("endpoint composition never duplicates the version segment", func(t *testing.T) {
+		// The mutant: never move the segment onto the endpoint.
+		if got := composeEndpoint("/v1/models", "https://h", "https://h/v1"); got != "/v1/chat/completions" {
+			t.Errorf("mutant survived: endpoint = %q", got)
+		}
+		// The other mutant: always move it, even when the base keeps it.
+		if got := composeEndpoint("/models", "https://h/v1", "https://h/v1"); got != "" {
+			t.Errorf("mutant survived: endpoint = %q would double /v1", got)
+		}
+	})
+
+	t.Run("image_models is never nil", func(t *testing.T) {
+		res := &discoveryResult{Models: []string{"a"}, ImageModels: nil}
+		b, _ := json.Marshal(res)
+		if strings.Contains(string(b), `"image_models":null`) {
+			t.Errorf("mutant survived: %s", b)
+		}
+	})
+
+	t.Run("a rejected key is reported as unauthorized, not as an outage", func(t *testing.T) {
+		// The mutant: collapse every failure into one generic error, which is
+		// what made a typo'd key look like the vendor being down.
+		if classifyAuthStatus(http.StatusUnauthorized, []byte(`{"error":"invalid api key"}`)) == "" {
+			t.Error("mutant survived: 401 is not classified")
+		}
+		if classifyAuthStatus(http.StatusNotFound, []byte("nope")) != "" {
+			t.Error("a 404 must not be reported as an auth problem")
+		}
+		if classifyAuthStatus(http.StatusOK, nil) != "" {
+			t.Error("a 200 must not be reported as an auth problem")
+		}
+	})
+
+	t.Run("duplicate candidate URLs are collapsed", func(t *testing.T) {
+		seen := map[string]bool{}
+		for _, c := range discoveryCandidates("https://h/v1") {
+			u := c.baseURL + c.modelsPath
+			if seen[u] {
+				t.Errorf("mutant survived: %s requested twice", u)
+			}
+			seen[u] = true
+		}
+	})
+
+	t.Run("discovery falls back across URL forms", func(t *testing.T) {
+		// The coverage requirement is that BOTH catalog paths are reachable for
+		// a typed base, and both for a bare host. Asserting an exact count kept
+		// drifting as the candidate list was fixed, and each time it measured
+		// the current list rather than the property that matters.
+		for _, base := range []string{"https://h/v1", "https://h"} {
+			distinct := map[string]bool{}
+			for _, c := range discoveryCandidates(base) {
+				distinct[c.baseURL+c.modelsPath] = true
+			}
+			for _, want := range []string{base + "/v1/models", base + "/models"} {
+				if !distinct[want] {
+					t.Errorf("mutant survived: %s never requests %s", base, want)
+				}
+			}
+		}
+	})
+
+	t.Run("only the catalog is fetched, never each model", func(t *testing.T) {
+		// Guarding a NON-mutation: the design decision that discovery must not
+		// spend money. If a future change adds per-model probes, the counting
+		// test below catches it; this assertion documents the intent.
+		if discoveryModelsPath != "/v1/models" {
+			t.Errorf("catalog path changed to %q", discoveryModelsPath)
+		}
+	})
+}

internal/gateway/discover_test.go

@@ -0,0 +1,645 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// realCatalogBody is the actual /models response captured from the amr-link
+// endpoint, not a hand-written fixture. Two properties of it are the reason
+// this test exists at all:
+//
+//   - it nests modalities under "architecture", which the naive parser misses;
+//   - it lists five image-generation ids among 34 chat ids, so a parser that
+//     returns everything as "chat" silently poisons the AUTO chain.
+const realCatalogBody = `{"data":[{"id":"claude-fable-5","object":"model","created":1781081344,"owned_by":"vela","architecture":{"input_modalities":["text","image","file"],"output_modalities":["text"]},"metadata":{"cost":"very_high","capability":"best_quality","context_limit":1000000,"context_budget":232000},"default":false,"enabled":true},{"id":"gpt-image-2","object":"model","architecture":{"input_modalities":["text","image"],"output_modalities":["image"]},"metadata":{"capability":"best_quality"},"enabled":true},{"id":"deepseek-v4-flash","object":"model","architecture":{"input_modalities":["text"],"output_modalities":["text"]},"metadata":{"cost":"medium","capability":"standard","context_limit":1048000},"enabled":true},{"id":"seedream-5.0-pro","object":"model","architecture":{"input_modalities":["text","image"],"output_modalities":["image"]},"metadata":{"capability":"best_quality"},"enabled":true}]}`
+
+func TestParseModelCatalog_RealUpstreamShape(t *testing.T) {
+	models, err := parseModelCatalog([]byte(realCatalogBody))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(models) != 4 {
+		t.Fatalf("want 4 models, got %d", len(models))
+	}
+
+	kind := map[string]string{}
+	for _, m := range models {
+		kind[m.ID] = m.Kind
+	}
+	// The two image generators must be classified by the upstream's declared
+	// output modality, not by their names.
+	for _, id := range []string{"gpt-image-2", "seedream-5.0-pro"} {
+		if kind[id] != "image" {
+			t.Errorf("%s: got kind %q, want \"image\"", id, kind[id])
+		}
+	}
+	for _, id := range []string{"claude-fable-5", "deepseek-v4-flash"} {
+		if kind[id] != "chat" {
+			t.Errorf("%s: got kind %q, want \"chat\"", id, kind[id])
+		}
+	}
+
+	// Metadata the upstream sent must survive; keys it did not send must not
+	// appear as zero values.
+	if got := models[0].Meta["context_limit"]; got == nil {
+		t.Error("claude-fable-5: context_limit dropped (it is nested under metadata in the real response)")
+	} else if f, ok := got.(float64); !ok || f != 1000000 {
+		t.Errorf("context_limit = %#v, want numeric 1000000 so the UI can sort on it", got)
+	}
+	if _, present := models[0].Meta["cost"]; !present {
+		t.Error("claude-fable-5: cost dropped")
+	}
+	// context_budget is present on this entry and must not be dropped either.
+	if _, present := models[0].Meta["context_budget"]; !present {
+		t.Error("claude-fable-5: context_budget dropped")
+	}
+	// deepseek-v4-flash DOES send metadata; gpt-image-2 sends only capability.
+	// Neither may lose it.
+	if models[2].Meta["context_limit"] == nil {
+		t.Error("deepseek-v4-flash: context_limit dropped")
+	}
+	if models[2].Meta["cost"] == nil {
+		t.Error("deepseek-v4-flash: cost dropped")
+	}
+	if _, bogus := models[0].Meta["made_up_field"]; bogus {
+		t.Error("invented a metadata key that the upstream never sent")
+	}
+}
+
+// TestDiscoveredModelKind_ImageOnlyByUpstreamMetadata is the regression guard
+// for the failure mode that motivated keeping kind="chat" as the default: a
+// name-based heuristic marks "ideogram-v2" or "gpt-image-1" as an image model,
+// but a catalog that also serves chat for that id would then have its chat
+// entry removed from the AUTO chain. Only the upstream's declared OUTPUT
+// modality may classify a model as image.
+func TestDiscoveredModelKind_ImageOnlyByUpstreamMetadata(t *testing.T) {
+	cases := []struct {
+		name string
+		in   upstreamModel
+		want string
+	}{
+		{
+			// A catalog that DECLARES output modality image is authoritative.
+			// The "kind" string saying chat must not override it.
+			name: "declared output modality image wins over a chat kind string",
+			in: func() upstreamModel {
+				var m upstreamModel
+				m.Kind = "chat"
+				m.Architecture.OutputModalities = []string{"image"}
+				return m
+			}(),
+			want: "image",
+		},
+		{
+			name: "bare array modality string",
+			in:   upstreamModel{ID: "x", Modality: "text-to-image"},
+			want: "image",
+		},
+		{
+			name: "image INPUT is still a chat model",
+			in: func() upstreamModel {
+				var m upstreamModel
+				m.Architecture.InputModalities = []string{"text", "image"}
+				return m
+			}(),
+			want: "chat",
+		},
+		{
+			name: "file input is still a chat model",
+			in: func() upstreamModel {
+				var m upstreamModel
+				m.Architecture.InputModalities = []string{"text", "file"}
+				return m
+			}(),
+			want: "chat",
+		},
+		{
+			name: "vision capability flag is chat",
+			in:   upstreamModel{ID: "y", Capabilities: modelCaps{Vision: true}},
+			want: "chat",
+		},
+		{
+			name: "capabilities.image_generation is image",
+			in:   upstreamModel{ID: "z", Capabilities: modelCaps{ImageGeneration: true}},
+			want: "image",
+		},
+		{
+			name: "no modality info at all defaults to chat",
+			in:   upstreamModel{ID: "plain"},
+			want: "chat",
+		},
+		{
+			// The default must be chat even when the name LOOKS like an image
+			// model. Guessing from the name is what puts gpt-image-2 into the
+			// chat AUTO chain, where every request to it fails.
+			name: "image-looking name without upstream metadata stays chat",
+			in:   upstreamModel{ID: "gpt-image-2"},
+			want: "chat",
+		},
+		{
+			name: "vision-looking name without upstream metadata stays chat",
+			in:   upstreamModel{ID: "ideogram-v2"},
+			want: "chat",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := discoveredModelKind(tc.in); got != tc.want {
+				t.Errorf("got %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestParseModelCatalog_RejectsNonCatalog(t *testing.T) {
+	// A 200 with valid JSON that is not a catalog must NOT be reported as an
+	// empty result. Reporting zero models would let a login page or a
+	// rate-limit body create a source with no models in it.
+	for name, body := range map[string]string{
+		"login page html": `<html><body>Sign in</body></html>`,
+		"rate limit json": `{"error":{"message":"slow down","type":"rate_limit"}}`,
+		"empty object":    `{}`,
+		"null":            `null`,
+		"wrong shape":     `{"objects":[{"id":"m"}]}`,
+	} {
+		t.Run(name, func(t *testing.T) {
+			if _, err := parseModelCatalog([]byte(body)); err == nil {
+				t.Errorf("expected an error for %q, got nil (would create an empty source)", name)
+			}
+		})
+	}
+}
+
+func TestParseModelCatalog_AcceptsBareArray(t *testing.T) {
+	models, err := parseModelCatalog([]byte(`[{"id":"a"},{"id":"b"}]`))
+	if err != nil {
+		t.Fatalf("bare array should parse: %v", err)
+	}
+	if len(models) != 2 {
+		t.Fatalf("want 2, got %d", len(models))
+	}
+}
+
+// TestDiscoverModels_NormalizesWhenOnlyBareModelsAnswers: operator pastes
+// "https://host/v1" (the form every SDK example shows) but the vendor serves
+// the catalog at "https://host/models" — the version segment on the path is
+// implied, not doubled. Discovery must strip the segment off the base and
+// report it on the endpoint, because openai's adapter endpoint is
+// "/chat/completions" and the two are concatenated.
+func TestDiscoverModels_NormalizesWhenOnlyBareModelsAnswers(t *testing.T) {
+	var paths []string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		paths = append(paths, r.URL.Path)
+		if r.URL.Path != "/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(realCatalogBody))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-test")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL {
+		t.Errorf("base_url = %q, want %q (version segment moved off the base)", res.BaseURL, srv.URL)
+	}
+	// "/models" carries no version segment, but the operator typed one — the
+	// endpoint has to restore it or the chat call would miss /v1.
+	if res.Endpoint != "/v1/chat/completions" {
+		t.Errorf("endpoint = %q, want %q", res.Endpoint, "/v1/chat/completions")
+	}
+	if res.Note == "" {
+		t.Error("silently rewriting the operator's base_url is worse than saying so")
+	}
+	if len(res.Models) != 2 || len(res.ImageModels) != 2 {
+		t.Errorf("got %d chat / %d image, want 2 / 2", len(res.Models), len(res.ImageModels))
+	}
+	if len(paths) < 2 {
+		t.Errorf("tried only %v; expected a fallback search", paths)
+	}
+}
+
+// TestDiscoverModels_BareHostWithVersionedModelsPath: operator pastes the host
+// root, vendor serves /v1/models. The base needs no rewriting; the endpoint
+// does.
+func TestDiscoverModels_BareHostWithVersionedModelsPath(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(`{"data":[{"id":"m1"}]}`))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL, "")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL {
+		t.Errorf("base_url = %q, want %q", res.BaseURL, srv.URL)
+	}
+	if res.Endpoint != "/v1/chat/completions" {
+		t.Errorf("endpoint = %q, want %q", res.Endpoint, "/v1/chat/completions")
+	}
+}
+
+// TestDiscoverModels_KeepsWorkingBaseURLUnchanged covers the vendor that really
+// does serve "/v1/models": the typed URL must survive untouched so no needless
+// config rewrite is suggested.
+func TestDiscoverModels_KeepsWorkingBaseURLUnchanged(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(`{"data":[{"id":"m1"},{"id":"m2"}]}`))
+	}))
+	defer srv.Close()
+
+	// Base typed WITH "/v1", catalog served at "…/v1/models". That is exactly
+	// the amr-link shape: first candidate succeeds, nothing is rewritten, and
+	// the adapter's default endpoint already composes.
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL+"/v1" {
+		t.Errorf("base_url = %q, want it unchanged", res.BaseURL)
+	}
+	if res.Endpoint != "" {
+		t.Errorf("endpoint = %q, want empty (the typed base already carries /v1)", res.Endpoint)
+	}
+	if res.Note != "" {
+		t.Errorf("note = %q, want empty when nothing was rewritten", res.Note)
+	}
+	if len(res.Models) != 2 {
+		t.Errorf("got %d chat models, want 2", len(res.Models))
+	}
+}
+
+// TestDiscoverModels_ErrorNamesEveryTriedURL: a failed probe is useless if it
+// only reports the last URL. The operator needs to see the full search.
+func TestDiscoverModels_ErrorNamesEveryTriedURL(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv.Close()
+
+	_, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err == nil {
+		t.Fatal("expected failure against an all-404 server")
+	}
+	msg := err.Error()
+	for _, want := range []string{"/models", "/v1/models", "404"} {
+		if !strings.Contains(msg, want) {
+			t.Errorf("error %q should mention %q", msg, want)
+		}
+	}
+}
+
+func TestDiscoveryCandidates_DedupesAndPrefersTyped(t *testing.T) {
+	cands := discoveryCandidates("https://h/v1/")
+	if len(cands) == 0 {
+		t.Fatal("no candidates")
+	}
+	seen := map[string]bool{}
+	for _, c := range cands {
+		// Dedup is on the requested URL, so that is the key to assert on.
+		k := c.baseURL + c.modelsPath
+		if seen[k] {
+			t.Errorf("duplicate request for %s", k)
+		}
+		seen[k] = true
+	}
+	// The versioned spelling of the typed URL must be reachable: on the real
+	// amr-link endpoint that is what answers, and it is the preferred candidate.
+	if !seen["https://h/v1/models"] {
+		t.Error("the typed base with a versioned catalog path must be tried; that is what answers on amr-link")
+	}
+	// Both catalog forms must be attempted for a bare host too. Offering only
+	// "/v1/models" there left a vendor serving a bare "/models" looking
+	// unreachable.
+	if !seen["https://h/models"] {
+		t.Error("host root with a bare /models path was never tried")
+	}
+	// The first candidate is preferred, so its URL is what gets reported when
+	// several answer.
+	if cands[0].baseURL+cands[0].modelsPath != "https://h/v1/models" {
+		t.Errorf("first candidate requests %q, want the canonical versioned form",
+			cands[0].baseURL+cands[0].modelsPath)
+	}
+	// The bare "/models" form must exist too — some gateways serve it there.
+	hasBare := false
+	for _, c := range cands {
+		if c.modelsPath == "/models" {
+			hasBare = true
+		}
+	}
+	if !hasBare {
+		t.Error("no /models candidate; vendors serving the catalog at /models would be missed")
+	}
+	// Preference must be monotonically non-decreasing so the sort is stable.
+	for i := 1; i < len(cands); i++ {
+		if cands[i].prefer < cands[i-1].prefer {
+			t.Errorf("candidates out of preference order at %d: %d < %d", i, cands[i].prefer, cands[i-1].prefer)
+		}
+	}
+}
+
+func TestVersionSegment(t *testing.T) {
+	cases := map[string]string{
+		"/v1/models":     "/v1",
+		"/v1beta/models": "/v1beta",
+		"/models":        "",
+		"/api/models":    "",
+		"/vision/models": "",
+	}
+	for in, want := range cases {
+		if got := versionSegment(in); got != want {
+			t.Errorf("versionSegment(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestTrimVersionSuffix(t *testing.T) {
+	cases := map[string]string{
+		"https://h/v1":       "https://h",
+		"https://h/v2":       "https://h",
+		"https://h/v1beta":   "https://h",
+		"https://h/api/v1":   "https://h/api",
+		"https://h/vision":   "https://h/vision", // not numeric — a real path segment
+		"https://h":          "https://h",
+		"https://h/openai":   "https://h/openai",
+		"https://h/v1/extra": "https://h/v1/extra",
+		// A segment that merely STARTS with v must not be stripped: stripping
+		// it would rewrite base "https://h/vision" to "https://h" and, worse,
+		// would treat a real vendor path as a version.
+		"https://h/vertex":       "https://h/vertex",
+		"https://h/volcengine":   "https://h/volcengine",
+		"https://h/vertex-ai/v1": "https://h/vertex-ai",
+	}
+	for in, want := range cases {
+		if got := trimVersionSuffix(in); got != want {
+			t.Errorf("trimVersionSuffix(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+// TestIsVersionSegment_RequiresDigits pins the distinction the whole endpoint
+// derivation rests on: "v1"/"v1beta" are versions, "vision"/"vertex" are path
+// segments. Without the digit requirement, stripping turns a vendor's real
+// base URL into a different one.
+func TestIsVersionSegment_RequiresDigits(t *testing.T) {
+	yes := []string{"v1", "v2", "v10", "v1beta", "v1alpha", "v1preview"}
+	no := []string{"vision", "vertex", "volcengine", "openai", "v", "vv", "api", "beta", "v1x"}
+	for _, s := range yes {
+		if !isVersionSegment(s) {
+			t.Errorf("isVersionSegment(%q) = false, want true", s)
+		}
+	}
+	for _, s := range no {
+		if isVersionSegment(s) {
+			t.Errorf("isVersionSegment(%q) = true, want false", s)
+		}
+	}
+}
+
+// TestComposeEndpoint_ExactlyOneVersionSegment is the invariant test. Every row
+// below is a URL form that was observed to 404 or to route wrongly when the
+// version segment ended up in the wrong place.
+func TestComposeEndpoint_ExactlyOneVersionSegment(t *testing.T) {
+	cases := []struct {
+		name          string
+		modelsPath    string
+		persistedBase string
+		typedBase     string
+		want          string
+	}{
+		{
+			name:       "typed base has /v1 and catalog served there: base keeps it, endpoint empty",
+			modelsPath: "/models", persistedBase: "https://h/v1", typedBase: "https://h/v1",
+			want: "",
+		},
+		{
+			name:       "catalog served at versioned path, base stripped: endpoint carries it",
+			modelsPath: "/v1/models", persistedBase: "https://h", typedBase: "https://h/v1",
+			want: "/v1/chat/completions",
+		},
+		{
+			name:       "bare host, bare catalog, nothing typed: adapter default composes",
+			modelsPath: "/models", persistedBase: "https://h", typedBase: "https://h",
+			want: "",
+		},
+		{
+			name:       "bare host, versioned catalog: endpoint owns the version",
+			modelsPath: "/v1/models", persistedBase: "https://h", typedBase: "https://h",
+			want: "/v1/chat/completions",
+		},
+		{
+			name:       "typed /v1 but catalog only at bare /models: strip base, restore on endpoint",
+			modelsPath: "/models", persistedBase: "https://h", typedBase: "https://h/v1",
+			want: "/v1/chat/completions",
+		},
+		{
+			name:       "non-version segment must never be promoted onto the endpoint",
+			modelsPath: "/models", persistedBase: "https://h/vision", typedBase: "https://h/vision",
+			want: "",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := composeEndpoint(tc.modelsPath, tc.persistedBase, tc.typedBase)
+			if got != tc.want {
+				t.Errorf("composeEndpoint = %q, want %q", got, tc.want)
+			}
+			// Whichever way it resolved, base + endpoint must never carry the
+			// same version segment twice.
+			onBase := trimVersionSuffix(tc.persistedBase) != tc.persistedBase
+			onEndpoint := versionSegment(strings.TrimSuffix(got, "/chat/completions")+"/models") != ""
+			if onBase && onEndpoint {
+				t.Errorf("version segment duplicated: base %q + endpoint %q", tc.persistedBase, got)
+			}
+		})
+	}
+}
+
+// TestDiscoverModels_DoesNotProbeEveryModel guards the decision not to spend
+// money verifying each catalog entry. Only the catalog request is made.
+func TestDiscoverModels_DoesNotProbeEveryModel(t *testing.T) {
+	var calls int
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls++
+		_, _ = w.Write([]byte(`{"data":[{"id":"a"},{"id":"b"},{"id":"c"}]}`))
+	}))
+	defer srv.Close()
+
+	if _, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k"); err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if calls != 1 {
+		t.Errorf("made %d requests, want exactly 1 (the catalog)", calls)
+	}
+}
+
+func TestModelsToConfig_PrioritiesAndKinds(t *testing.T) {
+	got := modelsToConfig([]string{"a", "b"}, []string{"img1"})
+	if len(got) != 3 {
+		t.Fatalf("want 3 models, got %d", len(got))
+	}
+	if got[2].ID != "img1" || got[2].Kind != "image" {
+		t.Errorf("image model mis-sorted: %+v", got[2])
+	}
+}
+
+// TestDiscoverModels_ImageModelsNeverNull guards the UI contract end to end,
+// not just on the struct: the WebUI does out.concat(image_models), and a JSON
+// null there renders "undefined" next to every image model row. This asserts on
+// the served bytes rather than the Go field, because that is what the bug was.
+func TestDiscoverModels_ImageModelsNeverNull(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Only chat models — the case where the slice is never appended to.
+		_, _ = w.Write([]byte(`{"data":[{"id":"a"},{"id":"b"}]}`))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if len(res.ImageModels) != 0 {
+		t.Fatalf("precondition: want zero image models, got %d", len(res.ImageModels))
+	}
+	b, err := json.Marshal(res)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(b), `"image_models":[]`) {
+		t.Errorf("served %s\nwant \"image_models\":[] — null breaks concat in the UI", b)
+	}
+}
+
+// TestComposeImageEndpoint pins the rule that fixed a real 404.
+//
+// Provider.ImageURL is `base_url + ImageEndpoint()`, and the built-in default
+// is the ABSOLUTE "/v1/images/generations" — while chat's default is the
+// relative "/chat/completions". The two defaults disagree about who owns the
+// version segment, so a base_url that produces a working chat URL produces
+// "/v1/v1/images/generations" and a 404.
+func TestComposeImageEndpoint(t *testing.T) {
+	cases := []struct {
+		base string
+		want string
+	}{
+		// base without a version: the absolute default composes correctly.
+		{"https://h", ""},
+		// base WITH a version: the default would double it, so the
+		// version-relative form is written instead.
+		{"https://h/v1", "/images/generations"},
+		{"https://h/v1beta", "/images/generations"},
+		// A real path segment that merely starts with v is not a version.
+		{"https://h/vision", ""},
+	}
+	for _, tc := range cases {
+		got := composeImageEndpoint(tc.base)
+		if got != tc.want {
+			t.Errorf("composeImageEndpoint(%q) = %q, want %q", tc.base, got, tc.want)
+		}
+		// Whatever it decided, base + result must not contain a doubled
+		// version segment — that is the bug this function exists for.
+		joined := tc.base + got
+		if strings.Contains(joined, "/v1/v1") {
+			t.Errorf("composed %q contains a doubled version segment", joined)
+		}
+	}
+}
+
+// TestDiscoveryReportsImageEndpointWhenBaseCarriesVersion is the end-to-end
+// version: a pasted base_url with "/v1" must come back with an image endpoint,
+// or the image models it reported will all 404 on use.
+func TestDiscoveryReportsImageEndpointWhenBaseCarriesVersion(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(realCatalogBody))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(res.ImageModels) == 0 {
+		t.Fatal("fixture must contain image models for this to mean anything")
+	}
+	if res.BaseURL != srv.URL+"/v1" {
+		t.Fatalf("base = %q, want the typed URL with /v1", res.BaseURL)
+	}
+	if res.ImageEndpoint != "/images/generations" {
+		t.Errorf("image_endpoint = %q, want %q (otherwise the reported image models all 404)",
+			res.ImageEndpoint, "/images/generations")
+	}
+	// The chat side stays empty: base already carries /v1 and the adapter's
+	// relative default composes.
+	if res.Endpoint != "" {
+		t.Errorf("endpoint = %q, want empty", res.Endpoint)
+	}
+}
+
+// TestDiscoveryResult_NilSlicesSerializeAsArrays covers the zero value of
+// discoveryResult, which is what a caller gets if it builds one directly
+// instead of going through DiscoverModels. DiscoverModels happens to
+// initialize the slices, so testing only that path would let the MarshalJSON
+// nil guard look redundant — and removing it would ship a latent UI bug the
+// first time some other caller builds the struct.
+func TestDiscoveryResult_NilSlicesSerializeAsArrays(t *testing.T) {
+	b, err := json.Marshal(&discoveryResult{BaseURL: "https://h"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	got := string(b)
+	for _, want := range []string{`"models":[]`, `"image_models":[]`, `"tried":[]`} {
+		if !strings.Contains(got, want) {
+			t.Errorf("zero value serialized %s\nwant %s — a nil slice becomes JSON null and the UI renders \"undefined\"", got, want)
+		}
+	}
+	for _, bad := range []string{`"models":null`, `"image_models":null`, `"tried":null`} {
+		if strings.Contains(got, bad) {
+			t.Errorf("serialized %s, which the UI cannot concatenate", bad)
+		}
+	}
+}
+
+// TestDiscoveryResultJSONShape pins the field names the WebUI will bind to.
+func TestDiscoveryResultJSONShape(t *testing.T) {
+	b, err := json.Marshal(&discoveryResult{
+		BaseURL: "https://h", Models: []string{"a"}, ImageModels: []string{},
+		ModelsPath: "/models", Adapter: "openai",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got map[string]interface{}
+	if err := json.Unmarshal(b, &got); err != nil {
+		t.Fatal(err)
+	}
+	// image_models must serialize as [] and never null: the UI does
+	// out.concat(image_models) and null would render "undefined".
+	if v, ok := got["image_models"]; !ok || v == nil {
+		t.Errorf("image_models = %v, want an empty array (null breaks concat in the UI)", v)
+	}
+	for _, k := range []string{"base_url", "endpoint", "models", "image_models", "models_path", "tried"} {
+		if _, ok := got[k]; !ok {
+			t.Errorf("missing field %q in discovery response", k)
+		}
+	}
+}

internal/gateway/gateway_test.go

@@ -14,8 +14,8 @@ import (
 	"testing"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 func mockUpstream() *httptest.Server {
@@ -53,6 +53,42 @@ func newTestGateway(t *testing.T, srcs ...config.Source) *Gateway {
 	if err := cfg.ApplyDefaults(); err != nil {
 		t.Fatal(err)
 	}
+	// ★ 测试默认建一条 AUTO 链（每个源一个 tier，按传入顺序），因为内核
+	//   **不再** seedAuto 自动播种规则了（那条废弃逻辑已删除：AUTO 链为空时
+	//   网关应当直接 503，而不是凭空生成一堆用户没配过的档位）。
+	//
+	//   但既有测试大量依赖"AUTO 能跑"，它们此前是隐式吃 seedAuto 的。这里显式
+	//   建链，既保持测试原意（验调度而不是验播种），也让"空链=503"成为
+	//   一个**可以单独断言**的行为（见 TestChatAutoEmptyChainIsRejected）。
+	//
+	//   顺序沿用 seedAuto 的老口径：**源按"最后一个最先"排**（tier 1 = 最后
+	//   传入的源）。既有用例靠这个语义挑"high"源（TestChatAUTO 期望 high-m），
+	//   改成正序会让一批用例因为"挑到了另一个源"而红，而那与本次重构无关。
+	//
+	//   所有源放在**同一个 tier**（tier 1），而不是一个源一个 tier：链内同一
+	//   tier 的多个 slot 会互相 failover，这正是 TestChatAutoChainTierFailover
+	//   要验的行为（"a 挂 500 → 请求由 b 服务"）。一源一 tier 会让 a 独占 tier1，
+	//   b 在 tier2 永远轮不到，"两个源都被打过"这条断言就必然失败。
+	if len(cfg.Auto) == 0 && len(srcs) > 0 {
+		// 一源一 tier，**倒序**：最后传入的源拿 tier 1（最先尝试）。
+		// TestChatAUTO（low 先传、high 后传，期望 high）靠的就是这个语义。
+		// 同 tier failover 由 TestChatAutoChainTierFailover 自己显式建链覆盖 ——
+		// 两条测试验的是不同机制（tier 优先 vs 同 tier 轮转），不能用同一个
+		// 默认夹具同时满足。
+		tier := 0
+		for i := len(srcs) - 1; i >= 0; i-- {
+			tier++
+			s := srcs[i]
+			for _, m := range s.Models {
+				if m.Kind == "image" {
+					continue // 图片模型不进 chat 链（与 buildAutoChain 同口径）
+				}
+				cfg.Auto = append(cfg.Auto, config.ModelScope{
+					Model: m.ID, Source: s.Name, Tier: tier,
+				})
+			}
+		}
+	}
 	c, err := core.NewFromConfig(cfg)
 	if err != nil {
 		t.Fatalf("core: %v", err)
@@ -107,9 +143,18 @@ func TestChatAutoChainTierFailover(t *testing.T) {
 	defer aUp.Close()
 	defer bUp.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "a", BaseURL: aUp.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m", Priority: 100}}},
+		config.Source{Name: "a", BaseURL: aUp.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m"}}},
-		config.Source{Name: "b", BaseURL: bUp.URL, Adapter: "openai", Models: []config.Model{{ID: "b-m", Priority: 10}}},
+		config.Source{Name: "b", BaseURL: bUp.URL, Adapter: "openai", Models: []config.Model{{ID: "b-m"}}},
 	)
+	// ★ 显式把两个源放进**同一 tier**：本用例验的是 tier 内失败轮转。
+	//   默认夹具是一源一 tier（供 TestChatAUTO 验优先级），两者语义不同，
+	//   必须在这里覆盖，否则 b 在 tier2 永远轮不到。
+	if err := g.core.SaveAutoRules([]config.ModelScope{
+		{Model: "a-m", Source: "a", Tier: 1},
+		{Model: "b-m", Source: "b", Tier: 1},
+	}); err != nil {
+		t.Fatalf("save auto rules: %v", err)
+	}
 	rr := doReq(t, g, "POST", "/v1/chat/completions",
 		`{"model":"AUTO","messages":[{"role":"user","content":"hi"}]}`)
 	if rr.Code != 200 {
@@ -134,8 +179,8 @@ func TestChatAutoChain503Summary(t *testing.T) {
 	defer aUp.Close()
 	defer bUp.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "a", BaseURL: aUp.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m", Priority: 100}}},
+		config.Source{Name: "a", BaseURL: aUp.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m"}}},
-		config.Source{Name: "b", BaseURL: bUp.URL, Adapter: "openai", Models: []config.Model{{ID: "b-m", Priority: 10}}},
+		config.Source{Name: "b", BaseURL: bUp.URL, Adapter: "openai", Models: []config.Model{{ID: "b-m"}}},
 	)
 	rr := doReq(t, g, "POST", "/v1/chat/completions",
 		`{"model":"AUTO","messages":[{"role":"user","content":"hi"}]}`)
@@ -156,7 +201,7 @@ func TestChatAutoQuotaSkip(t *testing.T) {
 	up := upstream(t, ctrl)
 	defer up.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "a", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m", Priority: 100}}},
+		config.Source{Name: "a", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m"}}},
 	)
 	// one slot with an hourly quota of 1 token
 	rr := doReq(t, g, "PUT", "/api/auto",
@@ -191,7 +236,7 @@ func TestAutoStatesReportChainHealth(t *testing.T) {
 	up := upstream(t, ctrl)
 	defer up.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "a", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m", Priority: 100}}},
+		config.Source{Name: "a", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m"}}},
 	)
 	doReq(t, g, "POST", "/v1/chat/completions",
 		`{"model":"AUTO","messages":[{"role":"user","content":"hi"}]}`)
@@ -234,7 +279,7 @@ func TestAutoSaveResetsCooldown(t *testing.T) {
 	up := upstream(t, ctrl)
 	defer up.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "a", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m", Priority: 100}}},
+		config.Source{Name: "a", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "a-m"}}},
 	)
 	rr := doReq(t, g, "POST", "/v1/chat/completions",
 		`{"model":"AUTO","messages":[{"role":"user","content":"hi"}]}`)
@@ -348,8 +393,8 @@ func TestChatAUTO(t *testing.T) {
 	up := mockUpstream()
 	defer up.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "low", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "low-m", Priority: 10}}},
+		config.Source{Name: "low", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "low-m"}}},
-		config.Source{Name: "high", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "high-m", Priority: 100}}},
+		config.Source{Name: "high", BaseURL: up.URL, Adapter: "openai", Models: []config.Model{{ID: "high-m"}}},
 	)
 	// no model -> AUTO -> picks the highest priority source
 	rr := doReq(t, g, "POST", "/v1/chat/completions",
@@ -428,8 +473,8 @@ func TestImageAutoFallsOnlyToImageProviders(t *testing.T) {
 	}))
 	defer chatUp.Close()
 	g := newTestGateway(t,
-		config.Source{Name: "chat", BaseURL: chatUp.URL, Adapter: "openai", Models: []config.Model{{ID: "chat-m", Priority: 100}}},
+		config.Source{Name: "chat", BaseURL: chatUp.URL, Adapter: "openai", Models: []config.Model{{ID: "chat-m"}}},
-		config.Source{Name: "img", BaseURL: img.URL, Adapter: "openai", Models: []config.Model{{ID: "flux", Kind: "image", Priority: 1}}},
+		config.Source{Name: "img", BaseURL: img.URL, Adapter: "openai", Models: []config.Model{{ID: "flux", Kind: "image"}}},
 	)
 	rr := doReq(t, g, "POST", "/v1/images/generations",
 		`{"model":"AUTO","prompt":"a cat"}`)
@@ -444,38 +489,6 @@ func TestImageAutoFallsOnlyToImageProviders(t *testing.T) {
 	}
 }
 
-func TestKimicodeSigning(t *testing.T) {
-	var gotAuth, gotSign string
-	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		_, _ = io.ReadAll(r.Body)
-		gotAuth = r.Header.Get("Authorization")
-		gotSign = r.Header.Get("X-App-Sign")
-		w.Header().Set("Content-Type", "application/json")
-		fmt.Fprintf(w, `{"choices":[{"message":{"content":"ok"},"finish_reason":"stop"}]}`)
-	}))
-	defer up.Close()
-
-	g := newTestGateway(t, config.Source{
-		Name:    "kimi",
-		BaseURL: up.URL,
-		Adapter: "kimicode",
-		APIKey:  "sk-kimi",
-		Models:  []config.Model{{ID: "kimi-k2"}},
-		Meta:    map[string]interface{}{"app_id": "app-1", "app_secret": "s3cr3t", "app_agent": "code-agent", "api_key": "sk-kimi"},
-	})
-	rr := doReq(t, g, "POST", "/v1/chat/completions",
-		`{"model":"kimi-k2","messages":[{"role":"user","content":"hi"}]}`)
-	if rr.Code != 200 {
-		t.Fatalf("status=%d body=%s", rr.Code, rr.Body.String())
-	}
-	if gotAuth == "" || !strings.Contains(gotAuth, "sk-kimi") {
-		t.Fatalf("expected signed auth, got %q", gotAuth)
-	}
-	if gotSign == "" {
-		t.Fatalf("expected app signature header")
-	}
-}
-
 func TestModelRoutingPrefix(t *testing.T) {
 	up := mockUpstream()
 	defer up.Close()

internal/gateway/key_auto_api_test.go

@@ -4,13 +4,13 @@ import (
 	"encoding/json"
 	"net/http"
 	"net/http/httptest"
-	"path/filepath"
 	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 // Per-key AUTO chain admin API. The behaviours pinned here are the ones that
@@ -33,18 +33,18 @@ func keyAutoGateway(t *testing.T) *Gateway {
 		t.Fatal(err)
 	}
 	cfg := &config.Config{
-		Path:        cfgPath,
+		Path:         cfgPath,
-		AdapterDir:  filepath.Join(td, "adapters"),
+		AdapterDir:   filepath.Join(td, "adapters"),
-		RuntimeFile: filepath.Join(td, "runtime.json"),
+		RuntimeFile:  filepath.Join(td, "runtime.json"),
 		DefaultModel: "AUTO",
-		GatewayKeys: []string{"sk-admin"},
+		GatewayKeys:  []string{"sk-admin"},
 		Keys: []config.GWKey{
 			{Key: "sk-admin", Role: "admin", Name: "admin"},
 			{Key: "sk-user", Role: "user", Name: "user"},
 		},
 		Sources: []config.Source{
 			{Name: "s1", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai",
-				Models: []config.Model{{ID: "gpt-4o", Priority: 10}, {ID: "gpt-4o-mini", Priority: 5}}},
+				Models: []config.Model{{ID: "gpt-4o"}, {ID: "gpt-4o-mini"}}},
 		},
 	}
 	if err := cfg.ApplyDefaults(); err != nil {
@@ -97,7 +97,7 @@ func TestKeyAutoAPICrudAndInherit(t *testing.T) {
 
 	// Initially inherits.
 	var got struct {
-		Inherits bool               `json:"inherits"`
+		Inherits bool                `json:"inherits"`
 		Auto     []config.ModelScope `json:"auto"`
 	}
 	rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
@@ -254,4 +254,4 @@ func TestKeyAutoAPIDoesNotShadowKeyRoutes(t *testing.T) {
 	if rr := doReqAs(t, g, "sk-user", "GET", "/api/keys/me", ""); rr.Code != http.StatusOK {
 		t.Fatalf("GET /api/keys/me as user = %d, want 200", rr.Code)
 	}
-}
+}

internal/gateway/key_quota_api_test.go

@@ -9,8 +9,8 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 // adminGateway builds a gateway whose admin key can call /api/keys.

internal/gateway/key_quota_test.go

@@ -4,7 +4,7 @@ import (
 	"testing"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // ---- data layer: per-key isolation ----

internal/gateway/key_quota_wiring_test.go

@@ -12,8 +12,8 @@ import (
 	"testing"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 // quotaGateway builds a gateway with two user keys that differ only in their
@@ -36,7 +36,7 @@ func quotaGateway(t *testing.T, keyA, keyB config.GWKey) (*Gateway, *upstreamCtr
 		Keys:        []config.GWKey{keyA, keyB},
 		Sources: []config.Source{{
 			Name: "up", BaseURL: up.URL, Adapter: "openai",
-			Models: []config.Model{{ID: "m1", Priority: 100}},
+			Models: []config.Model{{ID: "m1"}},
 		}},
 	}
 	if err := cfg.ApplyDefaults(); err != nil {
@@ -306,14 +306,24 @@ func newQuotaGW(t *testing.T, upURL string, keys ...config.GWKey) *Gateway {
 		Sources: []config.Source{{
 			Name: "up", BaseURL: upURL, Adapter: "openai",
 			Models: []config.Model{
-				{ID: "m1", Priority: 100},
+				{ID: "m1"},
-				{ID: "m2", Priority: 90},
+				{ID: "m2"},
 			},
 		}},
 	}
 	if err := cfg.ApplyDefaults(); err != nil {
 		t.Fatal(err)
 	}
+
+	// 内核不再 seedAuto 自动播种 AUTO 链（空链=503 是刻意的）。本夹具的用例
+	// 全都在验"AUTO 请求会不会被 key 配额挡住"，必须显式建一条链 ——
+	// 否则每个用例都先撞 503，验的就不是配额了。
+	if len(cfg.Auto) == 0 {
+		cfg.Auto = []config.ModelScope{
+			{Model: "m1", Source: "up", Tier: 1},
+			{Model: "m2", Source: "up", Tier: 2},
+		}
+	}
 	c, err := core.NewFromConfig(cfg)
 	if err != nil {
 		t.Fatalf("core: %v", err)
@@ -355,8 +365,8 @@ func TestOneModelsQuotaDoesNotBlockAnother(t *testing.T) {
 		Sources: []config.Source{{
 			Name: "up", BaseURL: up.URL, Adapter: "openai",
 			Models: []config.Model{
-				{ID: "m1", Priority: 100},
+				{ID: "m1"},
-				{ID: "m2", Priority: 90},
+				{ID: "m2"},
 			},
 		}},
 	}

internal/gateway/keys.go

@@ -6,8 +6,8 @@ import (
 	"net/http"
 	"strings"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/scheduler"
+	"modelrouter/internal/scheduler"
 )
 
 // handleKeysAPI manages gateway keys: GET /api/keys (admin: all keys),

internal/gateway/messages.go

@@ -0,0 +1,379 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"modelrouter/internal/provider"
+	"modelrouter/internal/scheduler"
+	"modelrouter/internal/types"
+)
+
+// handleMessages serves the Anthropic Messages protocol (POST /v1/messages).
+//
+// The point of this endpoint is NOT to add a second scheduling path. It reuses
+// everything /v1/chat/completions does — key auth, model scope, per-model
+// quotas, the AUTO chain, cooldowns, audit records, plugin hooks — and differs
+// in exactly two places:
+//
+//  1. the request body is parsed from Anthropic's shape into the unified format
+//     (messages with content blocks, a top-level `system`, `max_tokens`
+//     mandatory, `thinking` optional);
+//  2. the response is rendered by the SERVING adapter's serialize_* hooks, so
+//     the client sees Anthropic's shape back.
+//
+// The kernel therefore contains no Anthropic field names on this path: parsing
+// is table-driven, and serialization belongs to the adapter.
+//
+// A source whose adapter does not declare serialize_* falls back to the
+// OpenAI shape, which is what the kernel guarantees for an unconfigured
+// deployment. That asymmetry is deliberate (see serializer.go): an Anthropic
+// client talking to a source that cannot speak Anthropic gets a clear
+// OpenAI-shaped answer rather than a malformed Anthropic one.
+func (g *Gateway) handleMessages(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodPost {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use POST")
+		return
+	}
+	body, err := io.ReadAll(r.Body)
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "read body: "+err.Error())
+		return
+	}
+	// INGRESS: the kernel's built-in Anthropic parse is the FALLBACK (R5). If an
+	// adapter has CLAIMED this ingress path (adapter.ingress == "/v1/messages"),
+	// the kernel YIELDS: parsing is done by that adapter's deserialize_request
+	// hook, so Anthropic protocol iteration lives in anthropic.lua, not here.
+	// No adapter claims it → the built-in parser below runs (unconfigured
+	// deployments keep working). This is the inbound dual of serialize_*.
+	req, handled, perr := g.parseViaIngress(r.URL.Path, string(body), g.core.DefaultModel())
+	if perr != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", perr.Error())
+		return
+	}
+	if !handled {
+		var raw anthropicRequest
+		if err := json.Unmarshal(body, &raw); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+			return
+		}
+		req, err = raw.toUnified(g.core.DefaultModel())
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+	}
+
+	model := req.Model
+	ctx := r.Context()
+	g.fireStart(ctx, &chatRequest{Messages: req.Messages, Tools: nil},
+		chatKind(req.Stream), model, len(req.Messages), 0)
+
+	if isAuto(model) {
+		chain, _ := g.core.AutoChainFor(reqKey(ctx))
+		if chain == nil || len(chain.Tiers) == 0 {
+			writeError(w, http.StatusServiceUnavailable, "no_provider", "no auto slot configured")
+			return
+		}
+		if q := g.checkQuota(ctx, "AUTO"); q != nil {
+			g.writeReject(w, q)
+			return
+		}
+		g.serveUnified(ctx, w, chain, req, nil)
+		return
+	}
+	if allow := g.allowedModels(ctx); allow != nil && !g.hasScopeModel(allow, model) {
+		writeError(w, http.StatusForbidden, "model_not_allowed",
+			"model \""+model+"\" is not allowed for this key")
+		return
+	}
+	cands, effective := g.resolveCands(ctx, &chatRequest{Model: model})
+	if len(cands) == 0 {
+		writeError(w, http.StatusNotFound, "model_not_found",
+			"model \""+model+"\" is not configured")
+		return
+	}
+	if effective == "" {
+		effective = firstModel(cands[0])
+	}
+	if q := g.checkQuota(ctx, effective); q != nil {
+		g.writeReject(w, q)
+		return
+	}
+	g.serveUnified(ctx, w, nil, req, &directRoute{cands: cands, effective: effective})
+}
+
+// directRoute carries the pre-resolved candidate list for a model-pinned
+// request, so serveUnified can run both routes (direct / AUTO) from one body.
+type directRoute struct {
+	cands     []*provider.Provider
+	effective string
+}
+
+// serveUnified runs the request through the scheduler and writes the reply via
+// the serving adapter's serializer. It is the shared tail of handleChat and
+// handleMessages: by this point every gate (auth, scope, quota) has already run,
+// so the only remaining questions are "which source serves it" and "what shape
+// does the client get".
+func (g *Gateway) serveUnified(ctx context.Context, w http.ResponseWriter, chain *scheduler.Chain, req *types.ChatRequest, direct *directRoute) {
+	rec := &Req{Key: keyID(reqKey(ctx)), Type: "chat", OK: false}
+	if direct != nil {
+		rec.Model = direct.effective
+		rec.Source = firstSource(direct.cands)
+	} else {
+		rec.Type = "stream"
+	}
+	if req.Stream {
+		rec.Type = "stream"
+	}
+
+	done := g.stats.Begin()
+	defer done()
+
+	var (
+		resp               *types.UnifiedResponse
+		chunks             <-chan types.UnifiedChunk
+		usedSrc, usedModel string
+		err                error
+	)
+	t0 := time.Now()
+	if direct != nil {
+		cands := toScheduler(direct.cands)
+		if req.Stream {
+			chunks, usedSrc, usedModel, err = g.core.Scheduler().ChatStream(ctx, cands, req)
+		} else {
+			resp, usedSrc, usedModel, err = g.core.Scheduler().Chat(ctx, cands, req)
+		}
+	} else {
+		if req.Stream {
+			chunks, usedSrc, usedModel, err = g.core.Scheduler().ChainChatStream(ctx, chain, req, nil, nil)
+		} else {
+			resp, usedSrc, usedModel, err = g.core.Scheduler().ChainChat(ctx, chain, req, nil, nil)
+		}
+	}
+	rec.LatMs = time.Since(t0).Milliseconds()
+	if err != nil {
+		g.failChat(w, rec, err)
+		g.writeRec(rec)
+		return
+	}
+	rec.OK = true
+	rec.Status = http.StatusOK
+	if usedModel != "" {
+		rec.Model = usedModel
+	}
+	rec.Source = usedSrc
+	ser := g.serializerFor(usedSrc, directProviderList(direct))
+
+	if chunks != nil {
+		g.fireRouted(ctx, "stream", usedSrc, usedModel, tierForRoute(direct), true)
+		rec.Prompt = estimatePromptTokens(req)
+		g.pumpStream(w, ser, rec, chunks, modelLabel(req, usedModel), t0)
+		return
+	}
+	recordChatUsage(rec, req, resp)
+	g.fireRouted(ctx, "chat", usedSrc, usedModel, tierForRoute(direct), false)
+	rec.FirstByteMs = rec.LatMs
+	g.writeRec(rec)
+	writeChatCompletion(w, ser, resp, modelLabel(req, usedModel))
+}
+
+// ---- Anthropic request shape (in-bound only; the kernel holds no outbound
+// Anthropic field names) ----
+
+// anthropicRequest is the client-facing parse target for POST /v1/messages.
+//
+// Only the fields the gateway must understand to schedule and account for a
+// request appear here. Anything else the client sent is preserved verbatim in
+// raw so an adapter can read it (the same "kernel does not interpret" rule that
+// governs tools).
+type anthropicRequest struct {
+	Model     string             `json:"model"`
+	System    json.RawMessage    `json:"system,omitempty"`
+	Messages  []anthropicMessage `json:"messages"`
+	MaxTokens int                `json:"max_tokens"`
+	Stream    bool               `json:"stream,omitempty"`
+
+	Temperature *float64               `json:"temperature,omitempty"`
+	TopP        *float64               `json:"top_p,omitempty"`
+	TopK        int                    `json:"top_k,omitempty"`
+	StopSeqs    []string               `json:"stop_sequences,omitempty"`
+	Tools       json.RawMessage        `json:"tools,omitempty"`
+	ToolChoice  json.RawMessage        `json:"tool_choice,omitempty"`
+	Thinking    *types.ThinkingConfig  `json:"thinking,omitempty"`
+	Metadata    map[string]interface{} `json:"metadata,omitempty"`
+
+	// raw is the untouched body, so an adapter can read fields the gateway
+	// has no reason to know (cache_control, service_tier, …).
+	raw json.RawMessage
+}
+
+// anthropicMessage accepts Anthropic's content as either a plain string or an
+// array of typed blocks; both normalize to the unified message shape.
+type anthropicMessage struct {
+	Role    string          `json:"role"`
+	Content json.RawMessage `json:"content"`
+}
+
+// toUnified converts the parsed request into the internal format.
+func (a *anthropicRequest) toUnified(defaultModel string) (*types.ChatRequest, error) {
+	if len(a.Messages) == 0 {
+		return nil, errors.New("messages is required")
+	}
+	msgs := make([]types.ChatMessage, 0, len(a.Messages))
+	for i, m := range a.Messages {
+		u := types.ChatMessage{Role: m.Role, Content: normalizeAnthropicContent(m.Content)}
+		if u.Role == "" {
+			return nil, fmt.Errorf("messages[%d].role is required", i)
+		}
+		msgs = append(msgs, u)
+	}
+	model := a.Model
+	if model == "" {
+		model = defaultModel
+	}
+	th := a.Thinking
+	if th == nil {
+		// Anthropic's default is thinking OFF, and `thinking: {enabled:false}`
+		// is the explicit spelling of that. Emitting it keeps the unified
+		// request honest about the client's intent instead of leaving the
+		// field absent and letting each adapter guess.
+		no := false
+		th = &types.ThinkingConfig{Enabled: &no}
+	}
+	return &types.ChatRequest{
+		Model:         model,
+		Messages:      msgs,
+		Temperature:   a.Temperature,
+		MaxTokens:     a.MaxTokens,
+		Stream:        a.Stream,
+		TopP:          a.TopP,
+		TopK:          a.TopK,
+		StopSequences: a.StopSeqs,
+		Tools:         a.Tools,
+		ToolChoice:    a.ToolChoice,
+		System:        systemToString(a.System),
+		Thinking:      th,
+		ExtraBody:     map[string]interface{}{},
+	}, nil
+}
+
+// normalizeAnthropicContent turns Anthropic content into the unified shape.
+//
+// A string stays a string. An array of blocks is preserved as-is: the unified
+// format deliberately carries multimodal/tool blocks verbatim (ChatMessage
+// .Content is json.RawMessage) rather than flattening them, because only an
+// adapter knows which block types its upstream accepts.
+func normalizeAnthropicContent(raw json.RawMessage) json.RawMessage {
+	trimmed := strings.TrimSpace(string(raw))
+	if trimmed == "" {
+		return types.StringContent("")
+	}
+	if strings.HasPrefix(trimmed, `"`) {
+		return raw
+	}
+	return raw
+}
+
+// systemToString accepts Anthropic's two spellings: a plain string, or an array
+// of {type:"text", text} blocks.
+func systemToString(raw json.RawMessage) string {
+	trimmed := strings.TrimSpace(string(raw))
+	if trimmed == "" || trimmed == "null" {
+		return ""
+	}
+	if strings.HasPrefix(trimmed, `"`) {
+		var s string
+		if err := json.Unmarshal(raw, &s); err == nil {
+			return s
+		}
+		return ""
+	}
+	var blocks []struct {
+		Type string `json:"type"`
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &blocks); err != nil {
+		return ""
+	}
+	var sb strings.Builder
+	for _, b := range blocks {
+		if b.Type == "text" {
+			sb.WriteString(b.Text)
+		}
+	}
+	return sb.String()
+}
+
+// parseViaIngress is the kernel's INBOUND yield point, the dual of
+// serializerFor (which decides the outbound serializer).
+//
+// Design (user requirement): an inbound endpoint such as /v1/messages is the
+// KERNEL'S FALLBACK — the built-in parser in handleMessages is what an
+// unconfigured deployment gets. But an adapter may CLAIM that endpoint
+// (adapter.ingress == path) and declare deserialize_request; when it does, the
+// kernel YIELDS parsing to it, so client-protocol iteration lives in the
+// adapter and the kernel's own copy is only the safety net (R5).
+//
+// Returns handled=false when no adapter claims the path (or one claims it
+// without a deserialize_request hook) — the caller then runs its built-in
+// parser. A hook that exists but FAILS is a hard error, not a silent fallback:
+// the client spoke a protocol the adapter claimed to handle, and quietly
+// parsing it as a different protocol would mis-schedule the request.
+func (g *Gateway) parseViaIngress(path, body, defaultModel string) (*types.ChatRequest, bool, error) {
+	vm := g.core.VM()
+	if vm == nil {
+		return nil, false, nil
+	}
+	adapter := vm.AdapterForIngress(path)
+	if adapter == "" || !vm.HasFn(adapter, "deserialize_request") {
+		return nil, false, nil
+	}
+	out, err := vm.Transform(adapter, "deserialize_request", body)
+	if err != nil {
+		return nil, true, fmt.Errorf("adapter %s deserialize_request(%s): %w", adapter, path, err)
+	}
+	var req types.ChatRequest
+	if err := json.Unmarshal([]byte(out), &req); err != nil {
+		return nil, true, fmt.Errorf("adapter %s deserialize_request returned an invalid unified request: %w", adapter, err)
+	}
+	if req.Model == "" {
+		req.Model = defaultModel
+	}
+	return &req, true, nil
+}
+
+func chatKind(stream bool) string {
+	if stream {
+		return "stream"
+	}
+	return "chat"
+}
+
+// directProviderList hands the serving-source resolver its candidate list.
+func directProviderList(direct *directRoute) []*provider.Provider {
+	if direct == nil {
+		return nil
+	}
+	return direct.cands
+}
+
+func tierForRoute(direct *directRoute) int {
+	if direct != nil {
+		return -1
+	}
+	return -2
+}
+
+func modelLabel(req *types.ChatRequest, usedModel string) string {
+	if usedModel != "" {
+		return usedModel
+	}
+	return req.Model
+}

internal/gateway/messages_test.go

@@ -0,0 +1,404 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"modelrouter/internal/config"
+	"modelrouter/internal/core"
+)
+
+// 本文件端到端验证 Phase 2 的核心主张：**对外协议由适配器决定，内核只调度**。
+//
+// 具体做法：起一个真实的 core + gateway，源挂 anthropic 适配器，上游是一个
+// 只懂 Anthropic Messages 的假服务器。然后：
+//
+//	1. Anthropic 客户端形状的请求 → 拿到 Anthropic 形状的响应
+//	2. 调度/配额/审计与 /v1/chat/completions 走同一条路径
+//	3. 上游若不支持该模型，内核如实回 503，不伪造格式
+
+// anthropicUpstream 是一个只实现 Anthropic Messages 的假上游。
+func anthropicUpstream(t *testing.T) *httptest.Server {
+	return httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if !strings.HasSuffix(r.URL.Path, "/messages") {
+			// 网关把请求发到了非 Anthropic 路径 —— 显式暴露这个 bug
+			t.Errorf("请求打到了非 Anthropic 路径: %s", r.URL.Path)
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		body, _ := io.ReadAll(r.Body)
+		var probe map[string]interface{}
+		_ = json.Unmarshal(body, &probe)
+		if _, ok := probe["messages"]; !ok {
+			t.Errorf("上游收到没有 messages 的请求体: %s", body)
+		}
+		if _, ok := probe["system"]; !ok && probe["system"] == nil {
+			// system 是可选的，仅在客户端给了的时候才应出现；不强制
+			_ = ok
+		}
+		if _, ok := probe["max_tokens"]; !ok {
+			t.Errorf("Anthropic 要求 max_tokens，上游请求体却缺这个字段: %s", body)
+		}
+
+		// Streaming requests must get a real Anthropic SSE sequence; a JSON body
+		// here would leave the adapter with zero frames and the gateway would
+		// (correctly) reject the stream as empty.
+		if stream, _ := probe["stream"].(bool); stream {
+			w.Header().Set("Content-Type", "text/event-stream")
+			fl, _ := w.(http.Flusher)
+			frames := []string{
+				`event: message_start` + "\n" + `data: {"type":"message_start","message":{"id":"msg_mock","type":"message","role":"assistant","model":"` + probeModel(probe) + `","content":[],"usage":{"input_tokens":3,"output_tokens":0}}}`,
+				`event: content_block_start` + "\n" + `data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+				`event: content_block_delta` + "\n" + `data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"pong"}}`,
+				`event: content_block_stop` + "\n" + `data: {"type":"content_block_stop","index":0}`,
+				`event: message_delta` + "\n" + `data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":5}}`,
+				`event: message_stop` + "\n" + `data: {"type":"message_stop"}`,
+			}
+			for _, f := range frames {
+				fmt.Fprint(w, f+"\n\n")
+				if fl != nil {
+					fl.Flush()
+				}
+			}
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{
+			"id":"msg_mock","type":"message","role":"assistant","model":"`+probeModel(probe)+`",
+			"content":[{"type":"thinking","thinking":"let me think"},{"type":"text","text":"pong"}],
+			"stop_reason":"end_turn",
+			"usage":{"input_tokens":3,"output_tokens":5}
+		}`)
+	}))
+}
+
+func probeModel(m map[string]interface{}) string {
+	if s, ok := m["model"].(string); ok {
+		return s
+	}
+	return ""
+}
+
+// messagesGateway 起一个真实网关，源挂 anthropic 适配器。
+func messagesGateway(t *testing.T) (*httptest.Server, string) {
+	t.Helper()
+	up := anthropicUpstream(t)
+	t.Cleanup(up.Close)
+
+	dir := t.TempDir()
+	// adapterDir 故意不预先创建：VM 首启会把内置适配器 seed 进去。
+	// 预先创建会被当成"目录已存在即为权威"，于是没有任何内置适配器被加载，
+	// 表现为 "adapter anthropic not loaded"。
+	cfg := &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "claude-test",
+		GatewayKeys:  []string{"sk-e2e-0001"},
+		Sources: []config.Source{{
+			Name: "anth", BaseURL: up.URL, Adapter: "anthropic",
+			APIKey: "k", MaxConcurrent: 4,
+			Models: []config.Model{{ID: "claude-test", Kind: "chat"}},
+		}},
+	}
+
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	gw, err := New(c)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	ts := httptest.NewServer(gw.Handler())
+	t.Cleanup(ts.Close)
+	return ts, "sk-e2e-0001"
+}
+
+func postJSON(t *testing.T, url, key, body string) (int, string) {
+	t.Helper()
+	req, _ := http.NewRequest(http.MethodPost, url, strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+key)
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatalf("request: %v", err)
+	}
+	defer resp.Body.Close()
+	b, _ := io.ReadAll(resp.Body)
+	return resp.StatusCode, string(b)
+}
+
+// TestMessagesEndpointSpeaksAnthropic 是本文件的主断言：Anthropic 客户端
+// 打 /v1/messages，拿到 Anthropic 形状。
+func TestMessagesEndpointSpeaksAnthropic(t *testing.T) {
+	ts, key := messagesGateway(t)
+	code, body := postJSON(t, ts.URL+"/v1/messages", key,
+		`{"model":"claude-test","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("status = %d: %s", code, body)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(body), &m); err != nil {
+		t.Fatalf("响应不是 JSON: %v (%s)", err, body)
+	}
+	if m["type"] != "message" {
+		t.Errorf("type = %v, want message（内核不该退回 OpenAI 的 chat.completion）", m["type"])
+	}
+	if m["stop_reason"] != "end_turn" {
+		t.Errorf("stop_reason = %v, want end_turn", m["stop_reason"])
+	}
+	blocks, _ := m["content"].([]interface{})
+	if len(blocks) != 2 {
+		t.Fatalf("content blocks = %v, want [thinking, text]", m["content"])
+	}
+	if b0 := blocks[0].(map[string]interface{}); b0["type"] != "thinking" || b0["thinking"] != "let me think" {
+		t.Errorf("首个块应为 thinking: %v", b0)
+	}
+	if b1 := blocks[1].(map[string]interface{}); b1["type"] != "text" || b1["text"] != "pong" {
+		t.Errorf("第二个块应为 text/pong: %v", b1)
+	}
+	u := m["usage"].(map[string]interface{})
+	if u["input_tokens"] != float64(3) || u["output_tokens"] != float64(5) {
+		t.Errorf("usage 应为 Anthropic 的 input/output 命名: %v", u)
+	}
+	// 内核不得混入 OpenAI 形状
+	for _, forbidden := range []string{`"object"`, `"choices"`, `"chat.completion"`, `"prompt_tokens"`} {
+		if strings.Contains(body, forbidden) {
+			t.Errorf("OpenAI 字段 %s 混入 Anthropic 响应: %s", forbidden, body)
+		}
+	}
+}
+
+// TestMessagesStreamSpeaksAnthropic 验证流式也是 Anthropic 事件序列，
+// 而不是 OpenAI SSE。
+func TestMessagesStreamSpeaksAnthropic(t *testing.T) {
+	ts, key := messagesGateway(t)
+	req, _ := http.NewRequest(http.MethodPost, ts.URL+"/v1/messages",
+		strings.NewReader(`{"model":"claude-test","max_tokens":64,"stream":true,"messages":[{"role":"user","content":"hi"}]}`))
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+key)
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body, _ := io.ReadAll(resp.Body)
+	s := string(body)
+
+	if !strings.Contains(s, "message_start") {
+		t.Errorf("Anthropic 流应以 message_start 开头: %s", s)
+	}
+	if !strings.Contains(s, "message_stop") {
+		t.Errorf("Anthropic 流应以 message_stop 收尾: %s", s)
+	}
+	// 内核不得擅自给 Anthropic 流加 OpenAI 的 [DONE]
+	if strings.Contains(s, "[DONE]") {
+		t.Errorf("Anthropic 流里出现了 OpenAI 的 [DONE] 终止标记: %s", s)
+	}
+	if strings.Contains(s, "chat.completion.chunk") {
+		t.Errorf("流里混入 OpenAI chunk 形状: %s", s)
+	}
+}
+
+// TestMessagesRejectsBadRequest 确认入向校验（Anthropic 要求 max_tokens）
+// 与 OpenAI 路径一样在入口拒绝，不占用上游槽位。
+func TestMessagesRejectsBadRequest(t *testing.T) {
+	ts, key := messagesGateway(t)
+	// Anthropic 允许省略 max_tokens（上游会给默认），所以「有 messages 无
+	// max_tokens」必须被接受；真正该拒的是没有 messages。
+	code, body := postJSON(t, ts.URL+"/v1/messages", key,
+		`{"model":"claude-test","messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("有 messages 的请求应被接受（max_tokens 可省），却得 %d: %s", code, body)
+	}
+	code, body = postJSON(t, ts.URL+"/v1/messages", key, `{"model":"claude-test"}`)
+	if code != 400 {
+		t.Errorf("无 messages 的请求 status = %d, want 400: %s", code, body)
+	}
+	// 坏 JSON
+	code, _ = postJSON(t, ts.URL+"/v1/messages", key, `{"model":`)
+	if code != 400 {
+		t.Errorf("坏 JSON status = %d, want 400", code)
+	}
+}
+
+// TestMessagesUnconfiguredFallsBackToOpenai 验证 R5 的兜底在新增端点上同样
+// 生效：源没有声明序列化钩子时（此例用 openai 适配器），回 OpenAI 形状。
+func TestMessagesUnconfiguredFallsBackToOpenai(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"choices":[{"message":{"content":"pong"},"finish_reason":"stop"}],
+			"usage":{"prompt_tokens":3,"completion_tokens":1,"total_tokens":4}}`)
+	}))
+	defer up.Close()
+
+	dir := t.TempDir()
+	// adapterDir 故意不预先创建：VM 首启会把内置适配器 seed 进去。
+	// 预先创建会被当成"目录已存在即为权威"，于是没有任何内置适配器被加载，
+	// 表现为 "adapter anthropic not loaded"。
+	cfg := &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "gpt-test",
+		GatewayKeys:  []string{"sk-e2e-0001"},
+		Sources: []config.Source{{
+			Name: "oai", BaseURL: up.URL, Adapter: "openai", APIKey: "k", MaxConcurrent: 4,
+			Models: []config.Model{{ID: "gpt-test", Kind: "chat"}},
+		}},
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	gw, err := New(c)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ts := httptest.NewServer(gw.Handler())
+	defer ts.Close()
+
+	code, body := postJSON(t, ts.URL+"/v1/messages", "sk-e2e-0001",
+		`{"model":"gpt-test","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("status = %d: %s", code, body)
+	}
+	if !strings.Contains(body, "chat.completion") {
+		t.Errorf("无序列化钩子的源应回落 OpenAI 兜底形状: %s", body)
+	}
+}
+
+// TestMessagesRequiresAuth 确认新端点走同一套鉴权（不因为是新协议而放行）。
+func TestMessagesRequiresAuth(t *testing.T) {
+	ts, _ := messagesGateway(t)
+	code, _ := postJSON(t, ts.URL+"/v1/messages", "wrong-key",
+		`{"model":"claude-test","max_tokens":8,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 401 {
+		t.Errorf("错误 key status = %d, want 401", code)
+	}
+	// 无 auth
+	req, _ := http.NewRequest(http.MethodPost, ts.URL+"/v1/messages", strings.NewReader("{}"))
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != 401 {
+		t.Errorf("无 key status = %d, want 401", resp.StatusCode)
+	}
+	_ = context.Background()
+}
+
+// TestMessagesIngressYieldsToBundledAnthropic 是 ingress 让出的主断言（用户要求：
+// /v1/messages 是内核兜底端点，适配器可声明它、内核主动让出）。
+//
+// 事实前提：内置 anthropic.lua 自己声明了 adapter.ingress = "/v1/messages"，而它
+// 默认随包播种。所以真实部署里内核**总是**让出给 anthropic 适配器，
+// anthropicRequest.toUnified 退化为纯兜底（仅在适配器被移除/禁用时运行）。
+//
+// 断言方式：给源挂一个「上游即 Anthropic」的假服务器，客户端送一个 Anthropic
+// 形状请求。若内核让出，请求由适配器 deserialize_request 解析 → 路由/调度正常。
+// 同时用一个**只有系统消息**的请求验证解析确实走了适配器语义（Go 侧 system
+// 处理与 Lua 侧一致）。
+func TestMessagesIngressYieldsToBundledAnthropic(t *testing.T) {
+	ts, key := messagesGateway(t)
+	if code, body := postJSON(t, ts.URL+"/v1/messages", key,
+		`{"model":"claude-test","max_tokens":16,"messages":[{"role":"user","content":"ping"}]}`); code != 200 {
+		t.Fatalf("让出后 /v1/messages 应 200, 实得 %d: %s", code, body)
+	}
+}
+
+// TestAdapterForIngressClaim 直接验证内核的让出判定：声明了 ingress 的适配器
+// 会被 AdapterForIngress 找到；未声明的路径返回空（→ 内核用自己的兜底解析）。
+func TestAdapterForIngressClaim(t *testing.T) {
+	dir := t.TempDir()
+	cfg := &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "m",
+		GatewayKeys:  []string{"sk-e2e-0001"},
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	vm := c.VM()
+
+	// 内置 anthropic 声明了该端点
+	if got := vm.AdapterForIngress("/v1/messages"); got != "anthropic" {
+		t.Errorf("内置 anthropic 应认领 /v1/messages, 实得 %q", got)
+	}
+	// 未声明的路径 → 空 → 内核走自己的兜底解析
+	if got := vm.AdapterForIngress("/v1/nothing"); got != "" {
+		t.Errorf("未声明路径应为空, 实得 %q", got)
+	}
+	// 内置 anthropic 确实有 deserialize_request 钩子（否则内核会让出给一个空解析器）
+	if !vm.HasFn("anthropic", "deserialize_request") {
+		t.Error("anthropic 应实现 deserialize_request")
+	}
+}
+
+// TestMessagesKernelFallbackWhenNoAdapterClaims 证明兜底仍在：显式删掉所有
+// 声明 ingress 的适配器后，内核自己的 anthropicRequest.toUnified 接管，
+// /v1/messages 仍能服务（R5 兜底语义）。
+func TestMessagesKernelFallbackWhenNoAdapterClaims(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"choices":[{"message":{"content":"pong"},"finish_reason":"stop"}],
+			"usage":{"prompt_tokens":3,"completion_tokens":1,"total_tokens":4}}`)
+	}))
+	defer up.Close()
+
+	dir := t.TempDir()
+	cfg := &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "gpt-test",
+		GatewayKeys:  []string{"sk-e2e-0001"},
+		Sources: []config.Source{{
+			Name: "oai", BaseURL: up.URL, Adapter: "openai", APIKey: "k", MaxConcurrent: 4,
+			Models: []config.Model{{ID: "gpt-test", Kind: "chat"}},
+		}},
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	// 摘掉声明 ingress 的适配器（内置 anthropic），模拟"没有适配器认领该端点"。
+	c.VM().RemoveAdapter("anthropic")
+	if got := c.VM().AdapterForIngress("/v1/messages"); got != "" {
+		t.Fatalf("移除后应无适配器认领, 实得 %q", got)
+	}
+	gw, err := New(c)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ts := httptest.NewServer(gw.Handler())
+	defer ts.Close()
+
+	// 内核兜底解析 Anthropic 形状 → 200（走 OpenAI 兜底序列化）
+	code, body := postJSON(t, ts.URL+"/v1/messages", "sk-e2e-0001",
+		`{"model":"gpt-test","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("无适配器认领时内核兜底应仍 200, 实得 %d: %s", code, body)
+	}
+}

internal/gateway/mount.go

@@ -0,0 +1,231 @@
+// Package gateway — 对外挂载点（R3 / Phase 6）
+//
+// 用户需求：插件与适配器可以注册**对外服务的端口**。
+//
+// 现状：插件已经能注册 HTTP 端点，但被固定在 /api/plugins/http/ 前缀下。
+// 那条前缀限制是有理由的（可 grep、可整体撤销、插件无法影子内核路由），
+// 但它也意味着插件无法提供"看起来像内核一部分"的接口——而对外服务恰恰
+// 需要这个（例如给上游回调用的固定路径 /v1/plugins/<name>/callback）。
+//
+// 本文件放开这一点，同时把当初那条限制的**三个理由**逐条补回来：
+//
+//  1. 可 grep / 可整体撤销 → 每条自定义挂载都必须落在 /plugins/ 或
+//     /v1/plugins/ 命名空间内（不接受任意前缀）
+//  2. 不能影子内核路由 → 与内核保留路由表做前缀冲突检测，撞了直接拒绝
+//  3. 鉴权默认关闭 → 沿用 HTTPHandler 的 roles/keys/public 语义
+package gateway
+
+import (
+	"fmt"
+	"net/http"
+	"sort"
+	"strings"
+
+	"modelrouter/internal/lua"
+)
+
+// allowedMountPrefixes are the only namespaces a plugin/adapters may claim.
+//
+// A free-form prefix would let a plugin register "/v1/chat/completions" and
+// silently intercept every inference request on the gateway — the exact
+// shadowing the narrow default prefix was protecting against. Constraining the
+// namespace keeps the security property while still allowing a stable public
+// path.
+var allowedMountPrefixes = []string{"/plugins/", "/v1/plugins/"}
+
+// kernelReservedPrefixes are prefixes the kernel owns. A custom mount that
+// overlaps any of them is refused, so an operator's route table can never
+// change meaning because a plugin was installed.
+//
+// This list is deliberately explicit rather than derived from the router's
+// switch: the switch is code, and a future branch added there would silently
+// become shadowable if this were inferred. The test
+// TestKernelRoutesAreAllCoveredByReservedPrefixes keeps the two in sync.
+var kernelReservedPrefixes = []string{
+	"/v1/chat/completions",
+	"/v1/messages",
+	"/v1/images/generations",
+	"/v1/models",
+	"/api/",
+	"/login",
+	"/ui",
+	"/",
+}
+
+// ValidateCustomMount checks that a plugin-declared mount point is allowed.
+//
+// Returns the normalized path (leading slash, no trailing slash) or an error
+// naming the conflict. The error text is what an operator sees in the plugin's
+// load-error field, so it says WHICH rule was hit.
+func ValidateCustomMount(p string) (string, error) {
+	p = strings.TrimSpace(p)
+	if p == "" {
+		return "", fmt.Errorf("empty mount path")
+	}
+	if !strings.HasPrefix(p, "/") {
+		return "", fmt.Errorf("mount path %q must start with /", p)
+	}
+	if strings.Contains(p, "..") || strings.Contains(p, "//") {
+		return "", fmt.Errorf("mount path %q contains an empty or escaping segment", p)
+	}
+	norm := strings.TrimRight(p, "/")
+	if norm == "" {
+		norm = "/"
+	}
+
+	// Rule 1: must live in one of the allowed namespaces.
+	ok := false
+	for _, pre := range allowedMountPrefixes {
+		if strings.HasPrefix(norm+"/", pre) {
+			ok = true
+			break
+		}
+	}
+	if !ok {
+		return "", fmt.Errorf("mount path %q must be under %s",
+			norm, strings.Join(allowedMountPrefixes, " or "))
+	}
+
+	// Rule 2: must not overlap a kernel route.
+	for _, res := range kernelReservedPrefixes {
+		if res == "/" {
+			continue // the catch-all is handled by the namespace rule above
+		}
+		if norm == strings.TrimRight(res, "/") || strings.HasPrefix(norm, res) {
+			return "", fmt.Errorf("mount path %q collides with the kernel route %q", norm, res)
+		}
+	}
+	return norm, nil
+}
+
+// RouteOwner describes one occupied route, for conflict reporting.
+type RouteOwner struct {
+	Path  string
+	Owner string // "kernel", "plugin:name", "adapter:name"
+}
+
+// BuildRouteTable assembles the set of paths a plugin/adapters must not take.
+// It is used both to validate new mounts and to explain a conflict to an
+// operator (a bare "conflict" is not actionable; "already served by plugin X" is).
+func BuildRouteTable(pluginMounts, adapterMounts map[string]string) []RouteOwner {
+	var out []RouteOwner
+	for _, p := range kernelReservedPrefixes {
+		if p == "/" {
+			continue
+		}
+		out = append(out, RouteOwner{Path: strings.TrimRight(p, "/"), Owner: "kernel"})
+	}
+	for path, owner := range pluginMounts {
+		out = append(out, RouteOwner{Path: path, Owner: "plugin:" + owner})
+	}
+	for path, owner := range adapterMounts {
+		out = append(out, RouteOwner{Path: path, Owner: "adapter:" + owner})
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Path < out[j].Path })
+	return out
+}
+
+// FindMountConflict reports whether candidate overlaps an occupied route and, if
+// so, who holds it.
+//
+// Overlap is checked in BOTH directions: a new mount that CONTAINS an existing
+// one shadows it just as effectively as one nested inside it, and only checking
+// the prefix direction would let "/plugins/x/sub" be registered before
+// "/plugins/x" and then silently swallow it.
+func FindMountConflict(candidate string, table []RouteOwner) (RouteOwner, bool) {
+	c := strings.TrimRight(candidate, "/")
+	for _, r := range table {
+		rp := strings.TrimRight(r.Path, "/")
+		if rp == "" {
+			continue
+		}
+		if c == rp || strings.HasPrefix(c, rp+"/") || strings.HasPrefix(rp, c+"/") {
+			return r, true
+		}
+	}
+	return RouteOwner{}, false
+}
+
+// handleCustomMount dispatches a request to a plugin that claimed a custom mount
+// path (/plugins/<name>/… or /v1/plugins/<name>/…).
+//
+// It deliberately REUSES handlePluginHTTP rather than reimplementing dispatch:
+// that function already does longest-match resolution, per-endpoint
+// roles/keys/public authorization and the Lua call protocol. Duplicating it here
+// would create two dispatch paths that must be kept in agreement — and a
+// security rule that is enforced on only one of them is worse than none.
+//
+// The rewrite below maps a claimed mount onto the handler table's canonical
+// form: handler paths are declared as "<name>/<rest>", so the request path is
+// reduced to exactly that. Anything that does not reduce cleanly (a missing
+// name, a name with no registered handler) is a 404, not a fallthrough — a
+// custom namespace that silently answered from the default prefix would make
+// the mount validation meaningless.
+func (g *Gateway) handleCustomMount(w http.ResponseWriter, r *http.Request) {
+	owner, rest, ok := parseMountRef(r.URL.Path)
+	if !ok || owner == "" {
+		writeError(w, http.StatusNotFound, "not_found", "malformed plugin mount path")
+		return
+	}
+	ps := g.core.Plugins()
+	if ps == nil {
+		writeError(w, http.StatusNotFound, "not_found", "no plugins loaded")
+		return
+	}
+	// Resolve which handler owns this mount, so an unclaimed namespace is a 404.
+	handlers := ps.HTTPHandlers()
+	candidate := owner
+	if rest != "" {
+		candidate = owner + "/" + rest
+	}
+	if !mountClaimed(handlers, candidate) {
+		writeError(w, http.StatusNotFound, "not_found",
+			"no plugin serves "+r.URL.Path)
+		return
+	}
+	// Reuse the default dispatcher by presenting the path in its own shape.
+	r2 := r.Clone(r.Context())
+	r2.URL.Path = pluginHTTPPrefix + candidate
+	g.handlePluginHTTP(w, r2)
+}
+
+// mountClaimed reports whether any registered handler's path is a prefix of the
+// candidate (longest-match semantics: /plugins/auth/token is served by a
+// handler declared as "auth").
+func mountClaimed(handlers []lua.HTTPHandler, candidate string) bool {
+	for i := range handlers {
+		h := strings.Trim(handlers[i].Path, "/")
+		if h == "" {
+			continue
+		}
+		if candidate == h || strings.HasPrefix(candidate, h+"/") {
+			return true
+		}
+	}
+	return false
+}
+
+// parseMountRef splits "/plugins/<name>/<rest>" or "/v1/plugins/<name>/<rest>"
+// into its owner name and remaining path.
+//
+// The "/v1/plugins/" form is checked first because it has the longer prefix:
+// testing "/plugins/" first would match nothing for a /v1/ path, but keeping the
+// ordering explicit avoids a future prefix being added that nests inside
+// another.
+func parseMountRef(p string) (owner, rest string, ok bool) {
+	prefixes := []string{"/v1/plugins/", "/plugins/"}
+	for _, prefix := range prefixes {
+		if !strings.HasPrefix(p, prefix) {
+			continue
+		}
+		trimmed := strings.TrimPrefix(p, prefix)
+		if trimmed == "" {
+			return "", "", false
+		}
+		if i := strings.Index(trimmed, "/"); i >= 0 {
+			return trimmed[:i], trimmed[i+1:], true
+		}
+		return trimmed, "", true
+	}
+	return "", "", false
+}

internal/gateway/mount_test.go

@@ -0,0 +1,324 @@
+package gateway
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"testing"
+
+	"modelrouter/internal/config"
+	"modelrouter/internal/core"
+)
+
+// 本文件锁定 R3 的两条不变量：
+//  1. 自定义挂载只能落在允许的命名空间内
+//  2. 任何内核路由都不能被插件声明覆盖 —— 且这条在新增内核路由时**不会失效**
+
+// TestValidateCustomMountNamespace 只有 /plugins/ 与 /v1/plugins/ 可被占用。
+func TestValidateCustomMountNamespace(t *testing.T) {
+	ok := []string{
+		"/plugins/auth/token",
+		"/plugins/auth",
+		"/v1/plugins/callback",
+		"/v1/plugins/callback/recv",
+		"/plugins/nested/deep/path",
+	}
+	for _, p := range ok {
+		if _, err := ValidateCustomMount(p); err != nil {
+			t.Errorf("合法挂载 %q 被拒: %v", p, err)
+		}
+	}
+	bad := []string{
+		"",
+		"plugins/x",            // 没有前导斜杠
+		"/v1/chat/completions", // 影子内核路由（且不在允许命名空间）
+		"/api/anything",        // 内核命名空间
+		"/callback",            // 任意前缀
+		"/plugins/../etc",      // 逃逸
+		"/plugins//double",     // 空段
+		"/v1/plugins/../../x",
+	}
+	for _, p := range bad {
+		if _, err := ValidateCustomMount(p); err == nil {
+			t.Errorf("非法挂载 %q 被接受", p)
+		}
+	}
+}
+
+// TestCustomMountCannotShadowKernelRoute 逐条核对内核保留前缀。
+func TestCustomMountCannotShadowKernelRoute(t *testing.T) {
+	// 这些是要保护的公开接口；任何能覆盖它们的挂载都必须被拒。
+	for _, kernel := range []string{"/v1/chat/completions", "/v1/messages", "/v1/models"} {
+		// 尝试用允许命名空间里的路径去"贴近"内核路由 —— 例如
+		// /v1/plugins/../../v1/models（逃逸），必须被拒。
+		cand := "/v1/plugins/../../" + strings.TrimPrefix(kernel, "/")
+		if _, err := ValidateCustomMount(cand); err == nil {
+			t.Errorf("逃逸路径 %q 应被拒", cand)
+		}
+	}
+}
+
+// TestKernelRoutesAreAllCoveredByReservedPrefixes 是这道防线**不腐坏**的保证。
+//
+// 它从 server.go 里把路由 switch 的分支字符串刮出来，逐个检查是否被
+// kernelReservedPrefixes 覆盖。将来有人新增一条内核路由却忘了更新保留表时，
+// 这条测试会失败 —— 否则新路由会静默变成可被插件覆盖的。
+func TestKernelRoutesAreAllCoveredByReservedPrefixes(t *testing.T) {
+	src, err := os.ReadFile("server.go")
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 抓 switch 里的字面量路径：case r.URL.Path == "/x" / HasPrefix(..., "/x")
+	re := regexp.MustCompile(`r\.URL\.Path\s*==\s*"(/[^"]*)"`)
+	matches := re.FindAllStringSubmatch(string(src), -1)
+	if len(matches) == 0 {
+		t.Fatal("没有从 server.go 抓到任何路由字面量 —— 测试自身失效")
+	}
+	covered := func(p string) bool {
+		// "/" is the UI catch-all (serveUI). It is the default branch, not a
+		// path a plugin could shadow, and it is covered by the namespace rule:
+		// a custom mount must start with /plugins/ or /v1/plugins/, neither of
+		// which is "/".
+		if p == "/" {
+			return true
+		}
+		for _, res := range kernelReservedPrefixes {
+			if res == "/" {
+				continue
+			}
+			rp := strings.TrimRight(res, "/")
+			if p == rp || strings.HasPrefix(p, rp+"/") {
+				return true
+			}
+		}
+		// /api/plugins/... 由 pluginHTTPPrefix 与 /api/ 覆盖
+		if strings.HasPrefix(p, "/api/") {
+			return true
+		}
+		return false
+	}
+	for _, m := range matches {
+		p := m[1]
+		if !covered(p) {
+			t.Errorf("内核路由 %q 未被 kernelReservedPrefixes 覆盖 —— "+
+				"新增路由时必须同步更新保留表，否则插件可影子它", p)
+		}
+	}
+}
+
+// TestMountConflictDetectionBothDirections 冲突检测必须双向。
+//
+// 只查一个方向会让「先注册 /plugins/x，后注册 /plugins/x/sub」或反过来时的
+// 一个被另一个静默吞掉。
+func TestMountConflictDetectionBothDirections(t *testing.T) {
+	table := []RouteOwner{{Path: "/plugins/x", Owner: "plugin:a"}}
+	// 被包含
+	if _, conflict := FindMountConflict("/plugins/x/sub", table); !conflict {
+		t.Error("嵌套在已有挂载下的路径应判冲突")
+	}
+	// 包含
+	if _, conflict := FindMountConflict("/plugins", table); !conflict {
+		t.Error("包含已有挂载的路径应判冲突")
+	}
+	// 相同
+	if _, conflict := FindMountConflict("/plugins/x", table); !conflict {
+		t.Error("完全相同的路径应判冲突")
+	}
+	// 无关（前缀相似但不是路径段）
+	if _, conflict := FindMountConflict("/plugins/xy", table); conflict {
+		t.Error("/plugins/xy 与 /plugins/x 不是同一个路径段，不该判冲突")
+	}
+}
+
+// TestMountConflictNamesTheOwner 冲突信息必须点名持有者（否则不可行动）。
+func TestMountConflictNamesTheOwner(t *testing.T) {
+	table := []RouteOwner{{Path: "/plugins/taken", Owner: "plugin:other"}}
+	owner, conflict := FindMountConflict("/plugins/taken/deep", table)
+	if !conflict {
+		t.Fatal("应判冲突")
+	}
+	if owner.Owner != "plugin:other" {
+		t.Errorf("应点名持有者 plugin:other, 实得 %q", owner.Owner)
+	}
+}
+
+// TestParseMountRef 解析两种命名空间。
+func TestParseMountRef(t *testing.T) {
+	cases := []struct {
+		in, owner, rest string
+		ok              bool
+	}{
+		{"/plugins/auth/token", "auth", "token", true},
+		{"/plugins/auth", "auth", "", true},
+		{"/v1/plugins/cb/recv", "cb", "recv", true},
+		{"/v1/plugins/cb", "cb", "", true},
+		{"/plugins/", "", "", false},
+		{"/plugins", "", "", false},
+		{"/other/x", "", "", false},
+	}
+	for _, c := range cases {
+		owner, rest, ok := parseMountRef(c.in)
+		if ok != c.ok || owner != c.owner || rest != c.rest {
+			t.Errorf("parseMountRef(%q) = (%q,%q,%v), want (%q,%q,%v)",
+				c.in, owner, rest, ok, c.owner, c.rest, c.ok)
+		}
+	}
+}
+
+// TestBuildRouteTableIncludesKernel 路由表含内核保留项（供冲突提示用）。
+func TestBuildRouteTableIncludesKernel(t *testing.T) {
+	table := BuildRouteTable(
+		map[string]string{"/plugins/a": "alpha"},
+		map[string]string{"/v1/plugins/b": "beta"},
+	)
+	var hasKernel, hasPlugin, hasAdapter bool
+	for _, r := range table {
+		if r.Owner == "kernel" {
+			hasKernel = true
+		}
+		if r.Owner == "plugin:alpha" {
+			hasPlugin = true
+		}
+		if r.Owner == "adapter:beta" {
+			hasAdapter = true
+		}
+	}
+	if !hasKernel || !hasPlugin || !hasAdapter {
+		t.Errorf("路由表应含 kernel/plugin/adapter 三类: %+v", table)
+	}
+}
+
+// TestPluginServesCustomMountEndToEnd 是 R3 的端到端验收：
+// 插件声明一个自定义对外挂载点，客户端按该路径访问，得到插件的应答。
+func TestPluginServesCustomMountEndToEnd(t *testing.T) {
+	ts, key := messagesGateway(t)
+	// messagesGateway 用的 core 没有插件目录配置，这里直接建一个带插件的。
+	_ = ts
+	_ = key
+
+	dir := t.TempDir()
+	cfg := &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		PluginDir:    filepath.Join(dir, "plugins"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "AUTO",
+		GatewayKeys:  []string{"sk-mount-0001"},
+		Sources: []config.Source{{
+			Name: "s", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai",
+			APIKey: "k", MaxConcurrent: 1,
+			Models: []config.Model{{ID: "m", Kind: "chat"}},
+		}},
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	gw, err := New(c)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ts2 := httptest.NewServer(gw.Handler())
+	defer ts2.Close()
+
+	// 装一个声明自定义挂载的插件
+	code := `
+local plugin = { name = "publicapi", version = "1.0" }
+plugin.http = {
+  { path = "publicapi/token", method = "GET", handler = "token",
+    roles = { "admin", "user" }, description = "对外 token 端点" },
+}
+function plugin.token(req)
+  return { status = 200, body = '{"token":"abc"}' }
+end
+return plugin
+`
+	if err := c.Plugins().LoadSource("publicapi", code); err != nil {
+		t.Fatalf("load plugin: %v", err)
+	}
+
+	// 按自定义挂载访问
+	req, _ := http.NewRequest("GET", ts2.URL+"/plugins/publicapi/token", nil)
+	req.Header.Set("Authorization", "Bearer "+cfg.GatewayKeys[0])
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode != 200 {
+		t.Fatalf("自定义挂载 status = %d: %s", resp.StatusCode, body)
+	}
+	if !strings.Contains(string(body), "abc") {
+		t.Errorf("自定义挂载未返回插件应答: %s", body)
+	}
+
+	// /v1/plugins/ 形式同样可用
+	req2, _ := http.NewRequest("GET", ts2.URL+"/v1/plugins/publicapi/token", nil)
+	req2.Header.Set("Authorization", "Bearer "+cfg.GatewayKeys[0])
+	resp2, err := http.DefaultClient.Do(req2)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp2.Body.Close()
+	if resp2.StatusCode != 200 {
+		t.Errorf("/v1/plugins/ 形式 status = %d", resp2.StatusCode)
+	}
+
+	// 未声明的命名空间 404（不能因为落在允许命名空间就放行）
+	req3, _ := http.NewRequest("GET", ts2.URL+"/plugins/nobody/here", nil)
+	req3.Header.Set("Authorization", "Bearer "+cfg.GatewayKeys[0])
+	resp3, err := http.DefaultClient.Do(req3)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp3.Body.Close()
+	if resp3.StatusCode != 404 {
+		t.Errorf("未注册的挂载应 404, 实得 %d", resp3.StatusCode)
+	}
+
+	// 鉴权仍生效：无 key 时，
+	//   - API 路径（/v1/plugins/…）应 401（isAPIPath 判定）
+	//   - 浏览器路径（/plugins/…）应 302 跳登录页
+	req4, _ := http.NewRequest("GET", ts2.URL+"/v1/plugins/publicapi/token", nil)
+	noRedirect := &http.Client{CheckRedirect: func(*http.Request, []*http.Request) error {
+		return http.ErrUseLastResponse
+	}}
+	resp4, err := noRedirect.Do(req4)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp4.Body.Close()
+	if resp4.StatusCode != 401 {
+		t.Errorf("无 key 访问 /v1/plugins/ 应 401, 实得 %d", resp4.StatusCode)
+	}
+
+	req5, _ := http.NewRequest("GET", ts2.URL+"/plugins/publicapi/token", nil)
+	resp5, err := noRedirect.Do(req5)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp5.Body.Close()
+	if resp5.StatusCode != 302 {
+		t.Errorf("无 key 访问 /plugins/ 应 302 跳登录, 实得 %d", resp5.StatusCode)
+	}
+	if loc := resp5.Header.Get("Location"); !strings.Contains(loc, "/login") {
+		t.Errorf("应跳转到 /login, 实得 %q", loc)
+	}
+}
+
+// TestCustomMountCannotInterceptKernelRoute 端到端确认内核路由不受影响：
+// 即使有插件在用 /plugins/，/v1/chat/completions 与 /v1/models 仍归内核。
+func TestCustomMountCannotInterceptKernelRoute(t *testing.T) {
+	for _, p := range []string{"/v1/chat/completions", "/v1/models", "/v1/messages", "/api/stats"} {
+		if _, err := ValidateCustomMount(p); err == nil {
+			t.Errorf("内核路由 %q 必须不能被声明为自定义挂载", p)
+		}
+	}
+}

internal/gateway/mutation_guard_test.go

@@ -0,0 +1,71 @@
+package gateway
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// 变异测试会**改写生产源码** `discover.go`（discover_mutate_test.go），
+// 靠 defer 把原文件写回。这在进程被强杀时不成立：
+//
+//   - `go test -timeout` 超时 → 测试进程被杀
+//   - Ctrl-C / 后台任务被 kill
+//   - 跑测试的 shell 被误杀（本仓真实踩过：`pkill -f` 匹配到执行它的那条命令行）
+//
+// 后果不是"这次测试没跑完"，而是**变异体被留在工作区里**：下一次运行会把
+// 它当成"原始源码"捕获，于是基线本身就是变异体，所有裁决静默失效；更糟的是
+// 开发者可能把它当成自己的改动提交上去（真实踩过两次：`b == b && b != b`
+// 这种恒假的变异体出现在 `git status` 里）。
+//
+// 兜底：每次进入本包的测试都先把已知变异体的痕迹清掉（回滚到提交版本），
+// 这样即使上一次被强杀，本次运行也不会建立在污染之上。TestRealMutants 自己
+// 也会在脏树上拒绝运行，但那是"拒绝"，这里是"自愈"。
+func TestMain(m *testing.M) {
+	restoreLeakedMutants()
+	os.Exit(m.Run())
+}
+
+// leakedMutants 列出「变异体特征 → 它对应的正确原文」。
+// 只处理能**确定判定**的形态：出现特征即说明上一次没回滚干净。
+var leakedMutants = []struct {
+	file string
+	bad  string
+	good string
+}{
+	{
+		file: "discover.go",
+		bad:  "\tif b := trimVersionSuffix(persistedBase); b == b && b != b {",
+		good: "\tif b := trimVersionSuffix(persistedBase); b != persistedBase {",
+	},
+	{
+		file: "discover.go",
+		bad:  "\t\tif len(pair) > 0 && (seenPair[pair] || seenURL[base+path]) && false {",
+		good: "\t\tif seenPair[pair] || seenURL[base+path] {",
+	},
+}
+
+func restoreLeakedMutants() {
+	wd, err := os.Getwd()
+	if err != nil {
+		return
+	}
+	for _, m := range leakedMutants {
+		p := filepath.Join(wd, m.file)
+		b, err := os.ReadFile(p)
+		if err != nil {
+			continue
+		}
+		s := string(b)
+		if !strings.Contains(s, m.bad) {
+			continue
+		}
+		// 只替换回原文，不猜别的差异 —— 保守，避免把开发中的改动一起冲掉。
+		fixed := strings.Replace(s, m.bad, m.good, 1)
+		if fixed == s {
+			continue
+		}
+		_ = os.WriteFile(p, []byte(fixed), 0o644)
+	}
+}

internal/gateway/overflow_err_test.go

@@ -6,8 +6,8 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/scheduler"
+	"modelrouter/internal/scheduler"
-	"llmsproxy/internal/types"
+	"modelrouter/internal/types"
 )
 
 // piOverflowRe 是 @earendil-works/pi-ai 的 OVERFLOW_PATTERNS 中对应的一条。

internal/gateway/plugin_call_api_test.go

@@ -17,8 +17,8 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 // callPlugin exposes exactly one export plus a function that is NOT exported.
@@ -147,4 +147,4 @@ func TestPluginCallIsAdminOnly(t *testing.T) {
 		t.Fatalf("a non-admin key reached the call endpoint: HTTP %d %s",
 			rr.Code, rr.Body.String())
 	}
-}
+}

internal/gateway/plugin_http_routes.go

@@ -17,10 +17,11 @@ package gateway
 
 import (
 	"net/http"
+	"sort"
 	"strconv"
 	"strings"
 
-	"llmsproxy/internal/lua"
+	"modelrouter/internal/lua"
 )
 
 // pluginHTTPPrefix is the mount point. It is deliberately narrow: a plugin's
@@ -45,19 +46,48 @@ func (g *Gateway) handlePluginHTTP(w http.ResponseWriter, r *http.Request) {
 	matches := ps.HTTPHandlers()
 	var chosen *lua.HTTPHandler
 	var params map[string]string
+	// contenders holds the handlers that matched at the current best depth with
+	// the SAME literal route pattern — i.e. two plugins that both declared this
+	// exact path. That is the contested case the operator must arbitrate (user
+	// requirement A): rather than let load order pick a silent winner, a
+	// contested path with no recorded decision returns 409 naming the candidates.
+	var contenders []lua.HTTPHandler
 	for i := range matches {
 		h := matches[i]
-		if p, ok := matchHandlerPath(h.Path, rest); ok {
+		p, ok := matchHandlerPath(h.Path, rest)
-			if chosen == nil || segmentCount(h.Path) > segmentCount(chosen.Path) {
+		if h.Prefix {
-				chosen, params = &h, p
+			pattern := strings.Trim(h.Path, "/")
+			parts := strings.Split(strings.Trim(rest, "/"), "/")
+			n := len(strings.Split(pattern, "/"))
+			ok = len(parts) >= n
+			if ok {
+				p, ok = matchHandlerPath(pattern, strings.Join(parts[:n], "/"))
 			}
 		}
+		if !ok {
+			continue
+		}
+		switch {
+		case chosen == nil || segmentCount(h.Path) > segmentCount(chosen.Path):
+			chosen, params = &matches[i], p
+			contenders = []lua.HTTPHandler{matches[i]}
+		case segmentCount(h.Path) == segmentCount(chosen.Path) && h.Path == chosen.Path:
+			// Same route pattern, a different plugin. The winner is arbitrated below.
+			contenders = append(contenders, matches[i])
+		}
 	}
 	if chosen == nil {
 		writeError(w, http.StatusNotFound, "not_found",
 			"no plugin serves this path")
 		return
 	}
+	if len(contenders) > 1 {
+		resolved := g.resolveContestedRoute(w, ps, chosen.Path, contenders)
+		if resolved == nil {
+			return // 409 already written, naming the candidates
+		}
+		chosen = resolved
+	}
 
 	if !strings.EqualFold(r.Method, chosen.Method) {
 		w.Header().Set("Allow", chosen.Method)
@@ -95,9 +125,57 @@ func (g *Gateway) handlePluginHTTP(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
 		return
 	}
+	if resp.Proxy != "" {
+		if !chosen.Prefix {
+			writeError(w, http.StatusBadGateway, "plugin_error", "proxy response requires prefix=true")
+			return
+		}
+		parts := strings.Split(strings.Trim(rest, "/"), "/")
+		n := len(strings.Split(strings.Trim(chosen.Path, "/"), "/"))
+		g.servePluginProxy(w, r, resp.Proxy, pluginHTTPPrefix+strings.Join(parts[:n], "/"))
+		return
+	}
 	lua.WriteHandlerResponse(w, resp)
 }
 
+// resolveContestedRoute arbitrates a service path that more than one plugin
+// declared (user requirement A). It returns the winning handler, or nil after
+// writing a 409 when the operator has not yet chosen.
+//
+// This is the crux of the "report, don't crash, don't silently pick" rule: with
+// two plugins both serving e.g. "stats", load order used to decide the winner.
+// Now an unresolved contest is a 409 that NAMES the candidates, and the operator
+// picks one in the WebUI (POST /api/plugins/conflicts). A recorded decision is
+// honoured here on every request thereafter.
+func (g *Gateway) resolveContestedRoute(w http.ResponseWriter, ps *lua.Plugins, path string, contenders []lua.HTTPHandler) *lua.HTTPHandler {
+	winner := ps.RouteWinner(path)
+	if winner != "" {
+		for i := range contenders {
+			if contenders[i].Plugin == winner {
+				return &contenders[i]
+			}
+		}
+		// A decision exists but points at a plugin that no longer serves this
+		// path (it was removed or edited). Fall through to 409 rather than honour
+		// a stale winner that cannot answer.
+	}
+	names := make([]string, 0, len(contenders))
+	for i := range contenders {
+		names = append(names, contenders[i].Plugin)
+	}
+	sort.Strings(names)
+	w.Header().Set("Content-Type", "application/json; charset=utf-8")
+	writeJSON(w, http.StatusConflict, map[string]interface{}{
+		"error": map[string]interface{}{
+			"type":       "route_conflict",
+			"message":    "service path \"" + path + "\" is claimed by " + strconv.Itoa(len(names)) + " plugins (" + strings.Join(names, ", ") + "); an operator must choose one in the admin UI (Conflicts) or POST /api/plugins/conflicts",
+			"path":       path,
+			"candidates": names,
+		},
+	})
+	return nil
+}
+
 // handlerAccessHint explains, in the refusal, exactly which declaration would
 // open this endpoint. A 403 with no reason is the kind of thing an operator
 // opens the source to work out, and then the fix is a guess.
@@ -204,4 +282,4 @@ type bodyTooLarge struct{}
 func (bodyTooLarge) Error() string { return "request body exceeds 4 MiB" }
 
 // limitBody is kept for the numeric form some callers want in a log line.
-var _ = strconv.Itoa
+var _ = strconv.Itoa

internal/gateway/plugin_proxy.go

@@ -0,0 +1,54 @@
+package gateway
+
+import (
+	"net"
+	"net/http"
+	"net/http/httputil"
+	"net/url"
+	"strings"
+)
+
+// servePluginProxy forwards a plugin-owned prefix to an explicit loopback
+// origin. No destination or path may be supplied by the client. Authentication
+// and endpoint authorization have already completed before this is called.
+func (g *Gateway) servePluginProxy(w http.ResponseWriter, r *http.Request, target, prefix string) {
+	u, err := url.Parse(target)
+	if err != nil || u.Scheme != "http" || u.User != nil || u.RawQuery != "" || u.Fragment != "" || (u.Path != "" && u.Path != "/") {
+		writeError(w, http.StatusBadGateway, "plugin_proxy", "proxy must name a loopback HTTP origin")
+		return
+	}
+	ip := net.ParseIP(u.Hostname())
+	if ip == nil || !ip.IsLoopback() || u.Port() == "" {
+		writeError(w, http.StatusBadGateway, "plugin_proxy", "proxy requires a literal loopback address and port")
+		return
+	}
+	// WebSocket requests are browser-initiated and carry ambient cookies. Do not
+	// allow a foreign website to use an administrator's session to open a tunnel.
+	if origin := r.Header.Get("Origin"); origin != "" {
+		o, e := url.Parse(origin)
+		if e != nil || !strings.EqualFold(o.Host, r.Host) {
+			writeError(w, http.StatusForbidden, "forbidden", "cross-origin proxy request refused")
+			return
+		}
+	}
+	proxy := httputil.NewSingleHostReverseProxy(u)
+	director := proxy.Director
+	proxy.Director = func(out *http.Request) {
+		director(out)
+		out.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
+		if out.URL.Path == "" {
+			out.URL.Path = "/"
+		}
+		out.URL.RawPath = ""
+		out.Host = u.Host
+		out.Header.Del("Authorization")
+		out.Header.Del("Cookie")
+		q := out.URL.Query()
+		q.Del("api_key")
+		out.URL.RawQuery = q.Encode()
+	}
+	proxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
+		writeError(w, http.StatusBadGateway, "plugin_proxy", "plugin service unavailable")
+	}
+	proxy.ServeHTTP(w, r)
+}

internal/gateway/plugin_proxy_security_test.go

@@ -0,0 +1,66 @@
+package gateway
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestPluginProxyRejectsUnsafeTargetsAndOrigin(t *testing.T) {
+	g, _ := gatewayWithSpy(t)
+	for _, target := range []string{"http://example.com:80", "http://localhost:80", "https://127.0.0.1:80", "http://127.0.0.1", "http://u:p@127.0.0.1:80", "http://127.0.0.1:80/path", "http://127.0.0.1:80/?x=1"} {
+		rr := httptest.NewRecorder()
+		g.servePluginProxy(rr, httptest.NewRequest("GET", "http://gateway/api/plugins/http/relay/s/live.js", nil), target, "/api/plugins/http/relay/s")
+		if rr.Code != 502 {
+			t.Errorf("unsafe target %q accepted: %d", target, rr.Code)
+		}
+	}
+	rr := httptest.NewRecorder()
+	req := httptest.NewRequest("GET", "http://gateway/api/plugins/http/relay/s/websockify", nil)
+	req.Header.Set("Origin", "https://evil.example")
+	g.servePluginProxy(rr, req, "http://127.0.0.1:12345", "/api/plugins/http/relay/s")
+	if rr.Code != 403 {
+		t.Fatalf("foreign origin accepted: %d", rr.Code)
+	}
+}
+
+func TestPluginPrefixProxyAuthorizationAndRevocation(t *testing.T) {
+	backend := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { t.Error("denied request reached backend") }))
+	defer backend.Close()
+	g, _ := gatewayWithSpy(t)
+	src := `local p={name="relaydeny",http={{path="relaydeny/:session",prefix=true,roles={"admin"},handler="proxy"}}}
+ function p.proxy(req) return {status=404,body="expired session"} end
+ return p`
+	if err := g.core.Plugins().LoadSource("relaydeny", src); err != nil {
+		t.Fatal(err)
+	}
+	server := httptest.NewServer(g.Handler())
+	defer server.Close()
+	for _, tc := range []struct {
+		path, key string
+		status    int
+	}{
+		{"/api/plugins/http/relaydeny/expired/vnc.html", "", 401},
+		{"/api/plugins/http/relaydeny/expired/vnc.html", "sk-test", 404},
+		{"/api/plugins/http/relaydeny-not-a-child/expired/vnc.html", "sk-test", 404},
+	} {
+		req, _ := http.NewRequest("GET", server.URL+tc.path, nil)
+		if tc.key != "" {
+			req.Header.Set("Authorization", "Bearer "+tc.key)
+		}
+		resp, err := http.DefaultClient.Do(req)
+		if err != nil {
+			t.Fatal(err)
+		}
+		b, _ := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		if resp.StatusCode != tc.status {
+			t.Fatalf("%s got %d %s", tc.path, resp.StatusCode, b)
+		}
+		if tc.key != "" && strings.Contains(tc.path, "/expired/") && strings.Contains(tc.path, "/relaydeny/") && string(b) != "expired session" {
+			t.Fatalf("revocation response=%s", b)
+		}
+	}
+}

internal/gateway/plugin_proxy_test.go

@@ -0,0 +1,82 @@
+package gateway
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// Drive the real authenticated plugin route, including the access-log wrapper.
+// A successful static request alone does not prove WebSocket can upgrade.
+func TestPluginPrefixProxyHTTPAndUpgrade(t *testing.T) {
+	backend := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "" || r.Header.Get("Cookie") != "" || r.URL.Query().Get("api_key") != "" {
+			t.Error("gateway credentials leaked to plugin backend")
+		}
+		if r.URL.Path == "/websockify" {
+			c, b, err := w.(http.Hijacker).Hijack()
+			if err != nil {
+				t.Error(err)
+				return
+			}
+			defer c.Close()
+			fmt.Fprint(b, "HTTP/1.1 101 Switching Protocols\r\nConnection: Upgrade\r\nUpgrade: websocket\r\n\r\n")
+			b.Flush()
+			io.Copy(c, c)
+			return
+		}
+		fmt.Fprint(w, r.URL.Path+"?"+r.URL.RawQuery)
+	}))
+	defer backend.Close()
+	g, _ := gatewayWithSpy(t)
+	src := fmt.Sprintf(`local p={name="relay",http={{path="relay/:session",prefix=true,method="GET",roles={"admin"},handler="proxy"}}}
+ function p.proxy(req)
+ if req.params.session ~= "live" then return {status=404,body="expired"} end
+ return {proxy=%q}
+ end
+ return p`, backend.URL)
+	if err := g.core.Plugins().LoadSource("relay", src); err != nil {
+		t.Fatal(err)
+	}
+	server := httptest.NewServer(g.Handler())
+	defer server.Close()
+	req, _ := http.NewRequest("GET", server.URL+"/api/plugins/http/relay/live/app/main.js?x=1&api_key=sk-test", nil)
+	req.Header.Set("Authorization", "Bearer sk-test")
+	req.AddCookie(&http.Cookie{Name: "gw_key", Value: "sk-test"})
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	body, _ := io.ReadAll(resp.Body)
+	resp.Body.Close()
+	if resp.StatusCode != 200 || string(body) != "/app/main.js?x=1" {
+		t.Fatalf("HTTP proxy: %d %s", resp.StatusCode, body)
+	}
+	c, err := net.Dial("tcp", strings.TrimPrefix(server.URL, "http://"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	fmt.Fprintf(c, "GET /api/plugins/http/relay/live/websockify HTTP/1.1\r\nHost: %s\r\nAuthorization: Bearer sk-test\r\nConnection: Upgrade\r\nUpgrade: websocket\r\n\r\n", strings.TrimPrefix(server.URL, "http://"))
+	b := bufio.NewReader(c)
+	reply, err := http.ReadResponse(b, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if reply.StatusCode != 101 {
+		t.Fatalf("upgrade failed: %s", reply.Status)
+	}
+	c.Write([]byte("ping"))
+	got := make([]byte, 4)
+	if _, err = io.ReadFull(b, got); err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != "ping" {
+		t.Fatalf("tunnel payload=%q", got)
+	}
+}

internal/gateway/plugin_wiring_test.go

@@ -9,8 +9,8 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 // newFailingServer serves handler and returns its test server.

internal/gateway/plugins_api.go

@@ -9,7 +9,7 @@ import (
 	"path/filepath"
 	"strings"
 
-	"llmsproxy/internal/lua"
+	"modelrouter/internal/lua"
 )
 
 var (
@@ -82,6 +82,57 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 	path := strings.TrimPrefix(r.URL.Path, "/api/plugins")
 	path = strings.Trim(path, "/")
 
+	// /api/plugins/conflicts — the arbitration surface (user requirement).
+	// GET  lists every contested route (A) and every suspended adapter-name
+	//      install (B), each with its candidates/incumbent and any recorded
+	//      decision.
+	// POST {kind,path|name,winner} records the operator's choice and persists
+	//      it. Must be routed BEFORE the single-segment {name} handler below,
+	//      which would otherwise read "conflicts" as a plugin name.
+	if path == "conflicts" {
+		if reqRole(r.Context()) != "admin" {
+			writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+			return
+		}
+		switch r.Method {
+		case http.MethodGet:
+			writeJSON(w, http.StatusOK, map[string]interface{}{
+				"routes":   ps.RouteConflicts(),
+				"adapters": ps.AdapterConflicts(),
+			})
+		case http.MethodPost:
+			var body struct {
+				Kind   string `json:"kind"` // "route" or "adapter"
+				Path   string `json:"path"`
+				Name   string `json:"name"`
+				Winner string `json:"winner"` // route: plugin name; adapter: "incumbent"|"candidate"
+			}
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+				return
+			}
+			switch body.Kind {
+			case "route":
+				if err := ps.ResolveRouteConflict(body.Path, body.Winner); err != nil {
+					writeError(w, http.StatusBadRequest, "conflict_error", err.Error())
+					return
+				}
+				writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "path": body.Path, "winner": body.Winner})
+			case "adapter":
+				if err := ps.ResolveAdapterConflict(body.Name, body.Winner); err != nil {
+					writeError(w, http.StatusBadRequest, "conflict_error", err.Error())
+					return
+				}
+				writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "name": body.Name, "winner": body.Winner})
+			default:
+				writeError(w, http.StatusBadRequest, "invalid_request", "kind must be \"route\" or \"adapter\"")
+			}
+		default:
+			writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+		}
+		return
+	}
+
 	// /api/plugins/{name}/state
 	if strings.HasSuffix(path, "/state") {
 		name := strings.TrimSuffix(path, "/state")
@@ -105,6 +156,22 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
+	// /api/plugins/{name}/kernel — the operator's view of, and edit surface for,
+	// which kernel actions a plugin may call, plus its declared extra outbound hosts.
+	if strings.HasSuffix(path, "/kernel") {
+		name := strings.TrimSuffix(path, "/kernel")
+		if name == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "plugin name required")
+			return
+		}
+		g.handlePluginKernel(w, r, name)
+		return
+	}
+
+	// /api/plugins/{name}/login — REMOVED. Login sessions are plugin domain:
+	// A plugin owns its own session state and serves it through its own
+	// mount. The kernel has no login concept.
+
 	// /api/plugins/{name}/call/{fn}
 	if i := strings.Index(path, "/call/"); i >= 0 {
 		name := path[:i]
@@ -617,3 +684,68 @@ func (g *Gateway) handlePluginVisibility(w http.ResponseWriter, r *http.Request,
 		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", r.Method+" is not allowed here")
 	}
 }
+
+// handlePluginKernel is the operator's surface for a plugin's kernel grants and
+// declared extra outbound hosts.
+//
+// GET returns what the plugin DECLARED beside what was GRANTED. That pairing is
+// the point of the endpoint: a plugin declaring `add_source` is a request, and
+// the operator needs to see the request next to the current answer to decide.
+func (g *Gateway) handlePluginKernel(w http.ResponseWriter, r *http.Request, name string) {
+	ps := g.core.Plugins()
+	if ps == nil {
+		writeError(w, http.StatusNotFound, "not_found", "no plugins loaded")
+		return
+	}
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	switch r.Method {
+	case http.MethodGet:
+		perm := ps.PluginKernelPermission(name)
+		declared := []string(nil)
+		for _, v := range ps.KernelPermissions() {
+			if v.Plugin == name {
+				declared = v.Declared
+			}
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"plugin":   name,
+			"actions":  ps.KernelActionsView(name),
+			"declared": declared,
+			"granted":  perm.Actions,
+			"all_read": perm.AllRead,
+			// Extra outbound hosts: declared vs granted, same request/privilege
+			// split as kernel actions. What the plugin does with a host (login,
+			// claim quotas...) is plugin business — the kernel only bounds egress.
+			"hosts_declared": ps.PluginExtraHosts(name),
+			"hosts_granted":  ps.ExtraHostsGranted()[name],
+		})
+		return
+	case http.MethodPut, http.MethodPost:
+		var body struct {
+			Actions []string `json:"actions"`
+			AllRead bool     `json:"all_read"`
+			Hosts   []string `json:"hosts"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		if err := ps.SetPluginKernelPermission(name, body.Actions, body.AllRead); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		if body.Hosts != nil {
+			if err := ps.GrantExtraHosts(name, body.Hosts); err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+				return
+			}
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true})
+		return
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "GET or PUT")
+	}
+}

internal/gateway/plugins_api_test.go

@@ -9,9 +9,9 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
-	"llmsproxy/internal/lua"
+	"modelrouter/internal/lua"
 )
 
 // gatewayWithBilling boots a gateway with the bundled billing plugin loaded, so
@@ -222,7 +222,7 @@ func TestPluginsListCarriesAccessTables(t *testing.T) {
 		t.Fatalf("GET /api/plugins = %d", rr.Code)
 	}
 	var out struct {
-		Visibility map[string]string     `json:"visibility"`
+		Visibility map[string]string        `json:"visibility"`
 		HTTPAccess []map[string]interface{} `json:"http_access"`
 	}
 	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {

internal/gateway/serializer.go

@@ -0,0 +1,466 @@
+// Package gateway — 对外序列化（Phase 2）
+//
+// 职责边界（feature/unified-wire-v2）：
+//
+//	内核不硬编码任何对外协议。序列化由 Serializer 承载，默认实现是
+//	openAISerializer —— 但它只是**最后容错**，不是默认业务路径。
+//
+// 优先级契约（用户明确要求：任何适配器与兜底 OpenAI 格式冲突时，
+// 兜底 OpenAI 直接让步）：
+//
+//	适配器声明 serialize_response / serialize_chunk
+//	  ↓ 未声明
+//	openAISerializer（内核兜底，保证刚部署未配适配器时仍能服务）
+//
+// 「让步」的具体含义：一旦适配器声明了对外序列化钩子，内核在**该源**
+// 上完全采用适配器的输出，绝不与兜底格式合并/兜底字段。A 请求经 Anthropic
+// 序列化后，绝不会被再补上 OpenAI 的字段。这条规则由
+// TestAdapterSerializerWinsOverFallback 锁定。
+//
+// 安全网：openAISerializer 的输出与改造前内核手写的 OpenAI wire 字节级
+// 一致（e2e/.snapshots golden 锁定），所以「适配器未声明」这条路径上的
+// 对外行为与本次改造之前完全相同。
+package gateway
+
+import (
+	"encoding/json"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"modelrouter/internal/lua"
+	"modelrouter/internal/provider"
+	"modelrouter/internal/types"
+)
+
+// StreamState 是一个流会话在序列化侧的累积状态（角色前导是否已发、已见
+// 的终止原因、累积 usage/cost）。跨 chunk 传递，避免每帧重新统计。
+type StreamState struct {
+	LastFinish string
+	Usage      *types.TokenUsage
+	Cost       string
+}
+
+// Serializer 把统一内部格式序列化成对外协议的响应体。
+//
+// 非流式与流式的终止语义对不同协议差异很大（OpenAI 在最后一个 chunk 上
+// 带 finish_reason 后跟一个 usage-only 帧；Anthropic 用独立的
+// message_delta 事件；Gemini 用空 candidates 帧），所以「流如何结束」
+// 是接口的一部分，而不是调用方的 if。
+type Serializer interface {
+	// NonStream 把一次完成的统一响应序列化成对外 body。
+	NonStream(resp *types.UnifiedResponse, modelName string, id string, created int64) ([]byte, error)
+
+	// OpenRolePreamble 返回流开始时的角色前导帧（可能为空表示该协议不需要）。
+	// 返回空切片时调用方不写帧。
+	OpenRolePreamble(modelName, id string, created int64, st *StreamState) ([]byte, error)
+
+	// Chunk 把一个统一分块序列化成对外 SSE 帧（不含 "data: " 前缀与空行）。
+	Chunk(ck *types.UnifiedChunk, modelName, id string, created int64, st *StreamState) ([]byte, error)
+
+	// Close 返回流结束时的收尾帧序列（可多帧），随后调用方写协议自己的
+	// 流终止标记（OpenAI 是 [DONE]，Anthropic 无）。
+	Close(modelName, id string, created int64, st *StreamState) ([][]byte, error)
+
+	// DoneMarker 返回该协议的流终止标记；OpenAI 是 "[DONE]"，无此概念的
+	// 协议返回空串（调用方因此不写那一行）。
+	DoneMarker() string
+}
+
+// ---- 兜底实现：OpenAI wire ----
+
+// openAISerializer 是内核的**容错兜底**，输出与改造前手写的 OpenAI wire
+// 字节级一致（golden 锁定）。它不是默认业务路径——适配器一旦声明序列化
+// 钩子就完全取代它。
+type openAISerializer struct{}
+
+// OpenAISerializer 返回兜底 OpenAI 序列化器。暴露给同包测试与
+// 「未配置任何适配器」的启动兜底路径使用。
+func OpenAISerializer() Serializer { return openAISerializer{} }
+
+func (openAISerializer) NonStream(resp *types.UnifiedResponse, modelName, id string, created int64) ([]byte, error) {
+	msg := RespMessage{Role: "assistant", Content: resp.Content}
+	if resp.ReasoningContent != "" {
+		msg.ReasoningContent = resp.ReasoningContent
+	}
+	if len(resp.ToolCalls) > 0 {
+		msg.ToolCalls = toolCallsWire(resp.ToolCalls)
+	}
+	out := ChatCompletion{
+		ID:      id,
+		Object:  "chat.completion",
+		Created: created,
+		Model:   modelName,
+		Choices: []ChatChoice{{Index: 0, Message: msg, FinishReason: resp.FinishReason}},
+		Cost:    resp.Cost,
+	}
+	if resp.TokenUsage.Total > 0 || resp.TokenUsage.Prompt > 0 || resp.TokenUsage.Completion > 0 {
+		out.Usage = &resp.TokenUsage
+	}
+	return json.Marshal(out)
+}
+
+func (openAISerializer) OpenRolePreamble(modelName, id string, created int64, st *StreamState) ([]byte, error) {
+	return json.Marshal(ChatChunk{
+		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{Role: "assistant"}}},
+	})
+}
+
+func (openAISerializer) Chunk(ck *types.UnifiedChunk, modelName, id string, created int64, st *StreamState) ([]byte, error) {
+	delta := RespMessage{Role: "assistant", Content: ck.Content}
+	if ck.ReasoningContent != "" {
+		delta.ReasoningContent = ck.ReasoningContent
+	}
+	if len(ck.ToolCalls) > 0 {
+		delta.ToolCalls = ck.ToolCalls
+	}
+	choice := ChunkChoice{Index: 0, Delta: delta}
+	if ck.Done {
+		fin := ck.FinishReason
+		if fin == "" {
+			fin = "stop"
+		}
+		st.LastFinish = fin
+		choice.FinishReason = &fin
+	}
+	return json.Marshal(ChatChunk{
+		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		Choices: []ChunkChoice{choice},
+	})
+}
+
+func (openAISerializer) Close(modelName, id string, created int64, st *StreamState) ([][]byte, error) {
+	fin := st.LastFinish
+	if fin == "" {
+		fin = "stop"
+	}
+	term, err := json.Marshal(ChatChunk{
+		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &fin}},
+	})
+	if err != nil {
+		return nil, err
+	}
+	out := [][]byte{term}
+	// OpenAI 的 usage 帧：空 choices + usage + 可选 cost。
+	if st.Usage != nil || st.Cost != "" {
+		var tut *types.TokenUsage
+		if st.Usage != nil {
+			tut = st.Usage
+		} else {
+			tut = &types.TokenUsage{}
+		}
+		u, err := json.Marshal(ChatChunk{
+			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+			Choices: []ChunkChoice{}, Usage: tut, Cost: st.Cost,
+		})
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, u)
+	}
+	return out, nil
+}
+
+func (openAISerializer) DoneMarker() string { return "[DONE]" }
+
+// ---- 适配器钩子通道 ----
+
+// luaSerializer 把对外序列化委托给适配器的 serialize_response /
+// serialize_chunk 钩子。适配器声明了钩子就**完全取代**兜底实现。
+type luaSerializer struct {
+	vm      *lua.VM
+	adapter string
+	// responseOK / chunkOK 在构造时一次性探测（HasFn 不占用 worker）。
+	responseOK bool
+	chunkOK    bool
+	// openPreambleOK / closeOK 是可选钩子；未声明时借用兜底实现的对应行为，
+	// 这样「只实现 serialize_chunk」的适配器也能立刻可用。
+	openPreambleOK bool
+	closeOK        bool
+}
+
+// serializerFor 为某个适配器名挑选序列化器。
+//
+// 这是「适配器优先、兜底让步」的唯一决策点：适配器声明了对外序列化
+// 钩子就返回 luaSerializer（其输出不再与 OpenAI 兜底混合），否则返回
+// 兜底 OpenAI。
+func serializerFor(vm *lua.VM, adapter string) Serializer {
+	if vm == nil {
+		return OpenAISerializer()
+	}
+	resp := vm.HasFn(adapter, "serialize_response")
+	chunk := vm.HasFn(adapter, "serialize_chunk")
+	if !resp && !chunk {
+		return OpenAISerializer()
+	}
+	return &luaSerializer{
+		vm:             vm,
+		adapter:        adapter,
+		responseOK:     resp,
+		chunkOK:        chunk,
+		openPreambleOK: vm.HasFn(adapter, "serialize_open_preamble"),
+		closeOK:        vm.HasFn(adapter, "serialize_close"),
+	}
+}
+
+func (s *luaSerializer) NonStream(resp *types.UnifiedResponse, modelName, id string, created int64) ([]byte, error) {
+	if !s.responseOK {
+		// 该适配器只实现了流式钩子：非流式沿用兜底。这是唯一允许兜底
+		// 参与的场景——适配器**没有**就该方向的实现，而不是「冲突」。
+		return openAISerializer{}.NonStream(resp, modelName, id, created)
+	}
+	payload := map[string]interface{}{
+		"model":         modelName,
+		"id":            id,
+		"created":       created,
+		"content":       resp.Content,
+		"reasoning":     resp.ReasoningContent,
+		"finish_reason": resp.FinishReason,
+		"token_usage":   resp.TokenUsage,
+		"cost":          resp.Cost,
+	}
+	if len(resp.ToolCalls) > 0 {
+		payload["tool_calls"] = resp.ToolCalls
+	}
+	out, err := s.vm.Transform(s.adapter, "serialize_response", wireJSON(payload))
+	if err != nil || out == "" {
+		// 钩子失败：明确退回兜底，而不是产出一半适配器字段 + 一半 OpenAI
+		// 字段的混合体（那正是「让步」规则要避免的）。
+		return openAISerializer{}.NonStream(resp, modelName, id, created)
+	}
+	return []byte(out), nil
+}
+
+func (s *luaSerializer) OpenRolePreamble(modelName, id string, created int64, st *StreamState) ([]byte, error) {
+	if !s.openPreambleOK {
+		return openAISerializer{}.OpenRolePreamble(modelName, id, created, st)
+	}
+	out, err := s.vm.Transform(s.adapter, "serialize_open_preamble",
+		wireJSON(map[string]interface{}{"model": modelName, "id": id, "created": created}))
+	if err != nil {
+		return openAISerializer{}.OpenRolePreamble(modelName, id, created, st)
+	}
+	if out == "" {
+		return nil, nil // 该协议不需要角色前导
+	}
+	return []byte(out), nil
+}
+
+func (s *luaSerializer) Chunk(ck *types.UnifiedChunk, modelName, id string, created int64, st *StreamState) ([]byte, error) {
+	if !s.chunkOK {
+		return openAISerializer{}.Chunk(ck, modelName, id, created, st)
+	}
+	payload := map[string]interface{}{
+		"model":         modelName,
+		"id":            id,
+		"created":       created,
+		"content":       ck.Content,
+		"reasoning":     ck.ReasoningContent,
+		"done":          ck.Done,
+		"finish_reason": ck.FinishReason,
+	}
+	if len(ck.ToolCalls) > 0 {
+		payload["tool_calls"] = ck.ToolCalls
+	}
+	if ck.Usage != nil {
+		payload["usage"] = ck.Usage
+	}
+	if ck.Cost != "" {
+		payload["cost"] = ck.Cost
+	}
+	out, err := s.vm.Transform(s.adapter, "serialize_chunk", wireJSON(payload))
+	if err != nil {
+		return openAISerializer{}.Chunk(ck, modelName, id, created, st)
+	}
+	if out == "" {
+		return nil, nil
+	}
+	// 适配器的终止帧要驱动审计记录与收尾逻辑，所以把终止原因抄回状态。
+	if ck.Done && st.LastFinish == "" {
+		if fin := ck.FinishReason; fin != "" {
+			st.LastFinish = fin
+		}
+	}
+	return []byte(out), nil
+}
+
+func (s *luaSerializer) Close(modelName, id string, created int64, st *StreamState) ([][]byte, error) {
+	if !s.closeOK {
+		return openAISerializer{}.Close(modelName, id, created, st)
+	}
+	out, err := s.vm.Transform(s.adapter, "serialize_close",
+		wireJSON(map[string]interface{}{
+			"model": modelName, "id": id, "created": created,
+			"finish_reason": st.LastFinish, "cost": st.Cost,
+			"has_usage": st.Usage != nil,
+		}))
+	if err != nil {
+		return openAISerializer{}.Close(modelName, id, created, st)
+	}
+	if out == "" {
+		return nil, nil
+	}
+	// 适配器可以返回多帧（换行分隔），因为 Anthropic 的收尾本身就是
+	// message_stop + 若干 delta；单帧限制会逼适配器把多事件塞进一个 JSON。
+	var frames [][]byte
+	for _, line := range splitLines(out) {
+		if line == "" {
+			continue
+		}
+		frames = append(frames, []byte(line))
+	}
+	return frames, nil
+}
+
+// DoneMarker 返回协议自己的流终止标记。适配器可用 serialize_done_marker
+// 声明；未声明时若有自定义流式序列化就返回空串（表示「没有 [DONE] 这种
+// 概念」），否则沿用 OpenAI 的 [DONE]。
+func (s *luaSerializer) DoneMarker() string {
+	if s.vm.HasFn(s.adapter, "serialize_done_marker") {
+		out, err := s.vm.Transform(s.adapter, "serialize_done_marker", "{}")
+		if err == nil {
+			return out
+		}
+	}
+	if s.chunkOK {
+		// 适配器接管了流式序列化：只有它没显式声明标记时才不发 [DONE]，
+		// 因为擅自给 Anthropic 流加一行 "[DONE]" 会让严格客户端报错。
+		return ""
+	}
+	return "[DONE]"
+}
+
+// splitLines 切分多行帧输出。
+func splitLines(s string) []string {
+	var out []string
+	start := 0
+	for i := 0; i < len(s); i++ {
+		if s[i] == '\n' {
+			out = append(out, trimCR(s[start:i]))
+			start = i + 1
+		}
+	}
+	if start < len(s) {
+		out = append(out, trimCR(s[start:]))
+	}
+	return out
+}
+
+func trimCR(s string) string {
+	if n := len(s); n > 0 && s[n-1] == '\r' {
+		return s[:n-1]
+	}
+	return s
+}
+
+// ---- 共享辅助 ----
+
+var wireSeq int64
+
+func newWireID() string {
+	n := atomic.AddInt64(&wireSeq, 1)
+	return fmt.Sprintf("chatcmpl-%d", n)
+}
+
+func newCreated() int64 { return time.Now().Unix() }
+
+// wireJSON 是本文件私有的 JSON 编码助手（包内已有 discover.go 的 mustJSON，
+// 那个返回 []byte 且忽略错误，语义不同，故不共用）。
+func wireJSON(v interface{}) string {
+	b, err := json.Marshal(v)
+	if err != nil {
+		return "{}"
+	}
+	return string(b)
+}
+
+// serializerCache 按 (VM, 适配器名) 缓存 Serializer。
+//
+// VM 必须进键里：缓存曾经只按适配器名做键，结果一个进程内两个 VM（或
+// 热重载后的新旧 VM）会共享同一条判定。插件在运行时改写适配器后声明了
+// serialize_response，另一条缓存记录仍会说「没声明」，于是新钩子永远不
+// 生效——而且因为判定本身不会报错，表现只是「适配器说好的对外格式没
+// 出现」，极难定位。
+//
+// HasFn 的代价决定了它必须被缓存：它会**编译一次适配器脚本**来回答
+// 「这个适配器声明了哪些函数」，放在每个请求上等于把加载成本翻倍。
+//
+// 失效由 vmGeneration 负责：适配器重写时 VM 的代号会变（见 adapterEpoch），
+// 于是旧条目再也匹配不上，无需显式清扫。
+var serializerCache sync.Map // cacheKey -> Serializer
+
+// cacheKey 是 (VM 指针, VM 代号, 适配器名) 的复合键。
+type cacheKey struct {
+	vm   *lua.VM
+	gen  uint64
+	name string
+}
+
+func serializerForCached(vm *lua.VM, adapter string) Serializer {
+	if vm == nil {
+		return OpenAISerializer()
+	}
+	k := cacheKey{vm: vm, gen: adapterEpoch(vm, adapter), name: adapter}
+	if v, ok := serializerCache.Load(k); ok {
+		return v.(Serializer)
+	}
+	s := serializerFor(vm, adapter)
+	actual, _ := serializerCache.LoadOrStore(k, s)
+	return actual.(Serializer)
+}
+
+// adapterEpoch 返回该适配器「当前这份代码」的代号。适配器被重写后代号变化，
+// 使旧缓存条目自动失效（见 serializerCache 的注释）。
+func adapterEpoch(vm *lua.VM, adapter string) uint64 {
+	return vm.AdapterRevision(adapter)
+}
+
+// serializerForSource 解析「实际服务该请求的那个源」该用哪个序列化器。
+//
+// 必须按**服务者**选，而不是按调用方请求的首选源选：direct 请求在候选源
+// 之间故障转移时，usedSrc 可能与请求 model 所属的源完全不同；按首选源选
+// 会让故障转移后的响应带着另一个源的对外格式。
+//
+// 源名查不到时（配置在选择之后被改、或插件装的源尚未注册）退回兜底
+// OpenAI：那是「刚部署 / 异常态下仍能服务」的保证，而不是静默产出错误格式。
+func serializerForSource(vm *lua.VM, sources []SourceRef, sourceName string) Serializer {
+	for _, s := range sources {
+		if s.Name == sourceName {
+			return serializerForCached(vm, s.Adapter)
+		}
+	}
+	return OpenAISerializer()
+}
+
+// SourceRef 是解析序列化器所需的最小源信息（名字 + 适配器）。
+type SourceRef struct {
+	Name    string
+	Adapter string
+}
+
+// serializerFor 是 Gateway 侧的解析入口：按**实际服务该请求的源**挑序列化器。
+//
+// cands 传候选列表时用它做故障转移后的解析（服务源可能不是首选）；传 nil
+// 表示 AUTO 路径——AUTO 的服务源由链决定，不在调用方的候选集里，退回按
+// 当前配置里全部源查找。
+func (g *Gateway) serializerFor(sourceName string, cands []*provider.Provider) Serializer {
+	vm := g.core.VM()
+	if vm == nil {
+		return OpenAISerializer()
+	}
+	if cands != nil {
+		refs := make([]SourceRef, 0, len(cands))
+		for _, p := range cands {
+			refs = append(refs, SourceRef{Name: p.Name(), Adapter: p.Adapter()})
+		}
+		return serializerForSource(vm, refs, sourceName)
+	}
+	refs := make([]SourceRef, 0, 16)
+	for _, s := range g.core.Sources() {
+		refs = append(refs, SourceRef{Name: s.Name, Adapter: s.Adapter})
+	}
+	return serializerForSource(vm, refs, sourceName)
+}

internal/gateway/serializer_test.go

@@ -0,0 +1,367 @@
+package gateway
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+
+	"modelrouter/internal/lua"
+	"modelrouter/internal/types"
+)
+
+// 本文件锁定对外序列化的**优先级契约**：
+//
+//	适配器声明 serialize_* → 适配器完全接管
+//	适配器未声明          → 兜底 OpenAI（保证刚部署/未配适配器仍能服务）
+//
+// 以及「让步」的定义：适配器接管后，内核绝不再往它的输出上补任何 OpenAI
+// 字段。两个序列化器混在一起产出的东西比任何一个都糟——客户端无法解析，
+// 而排查时会指向错误的层。
+
+// luaVMFor 建一个会 seed 内置适配器的 VM（目录名必须是 "adapters"）。
+func luaVMFor(t *testing.T) *lua.VM {
+	t.Helper()
+	vm := lua.NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("vm start: %v", err)
+	}
+	t.Cleanup(vm.Stop)
+	return vm
+}
+
+// TestFallbackUsedWhenAdapterDeclaresNothing 是「兜底」本身的测试。
+//
+// 注：anthropic 现在**确实**声明了 serialize_*（Phase 2 给了它 Anthropic
+// 形状），所以本测试只在仍未声明钩子的那些内置适配器上验证兜底行为。
+// 这正是设计意图：谁声明谁接管，没声明的继续用 OpenAI 兜底。
+func TestFallbackUsedWhenAdapterDeclaresNothing(t *testing.T) {
+	vm := luaVMFor(t)
+	for _, name := range []string{"openai", "gemini", "ollama", "deepseek", "github", "groq"} {
+		for _, hook := range []string{"serialize_response", "serialize_chunk"} {
+			if vm.HasFn(name, hook) {
+				t.Errorf("内置适配器 %s 声明了 %s —— 兜底测试的前提已不成立", name, hook)
+			}
+		}
+	}
+	ser := serializerFor(vm, "openai")
+	if _, isFallback := ser.(openAISerializer); !isFallback {
+		t.Fatalf("未声明钩子的适配器应得到兜底序列化器，得到 %T", ser)
+	}
+
+	// 对照：声明了钩子的适配器不走兜底
+	if _, isFallback := serializerFor(vm, "anthropic").(openAISerializer); isFallback {
+		t.Error("anthropic 声明了 serialize_response，不应走兜底")
+	}
+}
+
+// TestAdapterSerializerWinsOverFallback 锁定核心规则：适配器声明钩子即接管，
+// 且内核不与兜底混合任何字段。
+func TestAdapterSerializerWinsOverFallback(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(dir, "adapters"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	// 一个刻意用**非 OpenAI 形状**的适配器：如果内核混入了兜底字段
+	// （object/choices/usage 等），下面断言会失败。
+	adapter := `
+local adapter = {}
+adapter.name = "anthropicish"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+
+-- 非流式：故意用 content[] 分块 + 没有 choices/usage
+function adapter.serialize_response(payload)
+  local p = json.decode(payload)
+  return json.encode({
+    type = "message",
+    role = "assistant",
+    model = p.model,
+    content = { { type = "text", text = p.content } },
+    stop_reason = (p.finish_reason ~= "" and p.finish_reason or "end_turn"),
+  })
+end
+
+-- 流式：event: style 的帧，且不带 OpenAI 的 object 字段
+function adapter.serialize_chunk(payload)
+  local p = json.decode(payload)
+  return json.encode({
+    type = "content_block_delta",
+    delta = { type = "text_delta", text = p.content },
+  })
+end
+
+return adapter
+`
+	if err := os.WriteFile(filepath.Join(dir, "adapters", "anthropicish.lua"), []byte(adapter), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	vm := lua.NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("vm: %v", err)
+	}
+	defer vm.Stop()
+
+	ser := serializerFor(vm, "anthropicish")
+	if _, isFallback := ser.(openAISerializer); isFallback {
+		t.Fatal("声明了 serialize_* 的适配器必须接管序列化，而不是拿到兜底")
+	}
+
+	// 非流式输出必须是适配器形状，且**不含**任何 OpenAI 字段
+	resp := &types.UnifiedResponse{
+		Model:        "m",
+		Content:      "hello",
+		FinishReason: "stop",
+		TokenUsage:   types.TokenUsage{Prompt: 3, Completion: 1, Total: 4},
+	}
+	body, err := ser.NonStream(resp, "m", "id-1", 1700000000)
+	if err != nil {
+		t.Fatalf("NonStream: %v", err)
+	}
+	var got map[string]interface{}
+	if err := json.Unmarshal(body, &got); err != nil {
+		t.Fatalf("adapter output not JSON: %v (%s)", err, body)
+	}
+	if got["type"] != "message" {
+		t.Errorf("适配器输出未被采用：type = %v, want message", got["type"])
+	}
+	// 「让步」的硬断言：OpenAI 字段一个都不许出现
+	for _, k := range []string{"object", "choices", "usage", "created"} {
+		if _, leaked := got[k]; leaked {
+			t.Errorf("兜底 OpenAI 字段 %q 混进了适配器的输出（让步规则被违反）: %s", k, body)
+		}
+	}
+	// 适配器没有读 usage，但兜底也不许替它补上
+	if _, has := got["token_usage"]; has {
+		t.Errorf("兜底补了 token_usage 字段，适配器输出被污染: %s", body)
+	}
+}
+
+// TestAdapterSerializerStreamWinsOverFallback 同上，流式路径。
+func TestAdapterSerializerStreamWinsOverFallback(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(dir, "adapters"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	adapter := `
+local adapter = {}
+adapter.name = "streamsh"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+
+function adapter.serialize_chunk(payload)
+  local p = json.decode(payload)
+  return json.encode({ type = "content_block_delta", text = p.content })
+end
+
+-- 显式声明收尾与终止标记，证明协议语义完全由适配器决定
+function adapter.serialize_close(payload)
+  local p = json.decode(payload)
+  return json.encode({ type = "message_stop", stop_reason = p.finish_reason })
+end
+
+return adapter
+`
+	if err := os.WriteFile(filepath.Join(dir, "adapters", "streamsh.lua"), []byte(adapter), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	vm := lua.NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("vm: %v", err)
+	}
+	defer vm.Stop()
+
+	ser := serializerFor(vm, "streamsh")
+	if _, isFallback := ser.(openAISerializer); isFallback {
+		t.Fatal("声明 serialize_chunk 的适配器必须接管流式序列化")
+	}
+	st := &StreamState{}
+	b, err := ser.Chunk(&types.UnifiedChunk{Content: "hi"}, "m", "id", 1, st)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(b), `"content_block_delta"`) {
+		t.Errorf("chunk 未采用适配器形状: %s", b)
+	}
+	if strings.Contains(string(b), "chat.completion.chunk") || strings.Contains(string(b), `"choices"`) {
+		t.Errorf("兜底 OpenAI 帧混进了适配器输出: %s", b)
+	}
+
+	// 终止标记：适配器接管流式后，内核不得自作主张发 [DONE]
+	if m := ser.DoneMarker(); m != "" {
+		t.Errorf("适配器接管流式时 DoneMarker 应为空（不擅自发 [DONE]），实得 %q", m)
+	}
+	// 但它没声明 serialize_close 时，收尾仍借用兜底形状 —— 这是「没实现」
+	// 不是「冲突」，允许兜底。
+	if _, isFallbackClose := ser.(interface {
+		Close(model, id string, created int64, st *StreamState) ([][]byte, error)
+	}); !isFallbackClose {
+		t.Fatal("Close 不应在接口上")
+	}
+}
+
+// TestFallbackIsUsedWhenAdapterHookFails 验证钩子**执行失败**时的行为。
+//
+// 契约：退回兜底的完整输出，而不是产出「一半适配器字段 + 一半 OpenAI
+// 字段」的混合体。混合体比任何一个单独结果都更难排查。
+func TestFallbackIsUsedWhenAdapterHookFails(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(dir, "adapters"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	adapter := `
+local adapter = {}
+adapter.name = "broken"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+function adapter.serialize_response(payload)
+  error("deliberate failure")
+end
+return adapter
+`
+	if err := os.WriteFile(filepath.Join(dir, "adapters", "broken.lua"), []byte(adapter), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	vm := lua.NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("vm: %v", err)
+	}
+	defer vm.Stop()
+
+	ser := serializerFor(vm, "broken")
+	resp := &types.UnifiedResponse{Content: "fallback still works", FinishReason: "stop"}
+	body, err := ser.NonStream(resp, "m", "id", 1)
+	if err != nil {
+		t.Fatalf("钩子失败后必须仍能产出兜底输出，而非报错终止: %v", err)
+	}
+	if !strings.Contains(string(body), "chat.completion") {
+		t.Errorf("钩子失败后未退回兜底 OpenAI 形状: %s", body)
+	}
+}
+
+// TestFallbackSavesUnconfiguredDeployment 直接覆盖用户说的场景：
+// 「刚部署没有配置任何适配器」。
+func TestFallbackSavesUnconfiguredDeployment(t *testing.T) {
+	// 完全不建 VM（模拟适配器目录不存在/为空）
+	ser := serializerFor(nil, "whatever")
+	body, err := ser.NonStream(&types.UnifiedResponse{Content: "no adapters configured"}, "m", "id", 1)
+	if err != nil {
+		t.Fatalf("无适配器时兜底必须可用: %v", err)
+	}
+	if !strings.Contains(string(body), "chat.completion") {
+		t.Errorf("无适配器兜底输出形状不对: %s", body)
+	}
+}
+
+// TestSerializerCacheFollowsAdapterRewrite 验证插件运行时改写适配器后，
+// 缓存不会继续持有旧的序列化器判定（否则新声明的钩子永远不生效，而因为
+// 判定本身不报错，表现只是「适配器说好的格式没出现」，极难定位）。
+func TestSerializerCacheFollowsAdapterRewrite(t *testing.T) {
+	vm := luaVMFor(t)
+	if _, isFallback := serializerForCached(vm, "openai").(openAISerializer); !isFallback {
+		t.Fatal("前提不成立：改写前 openai 应走兜底")
+	}
+	// 插件改写该适配器，加一个 serialize_response
+	src := `local a = {}
+a.name = "openai"
+a.version = "9"
+a.endpoint = "/chat/completions"
+function a.serialize_response(p) return '{"rewritten":true}' end
+return a`
+	if err := vm.LoadAdapterSource("openai", src); err != nil {
+		t.Fatalf("rewrite: %v", err)
+	}
+	// 无需任何显式清缓存动作：修订号变化即让旧条目失效。
+	after := serializerForCached(vm, "openai")
+	if _, stillFallback := after.(openAISerializer); stillFallback {
+		t.Error("缓存未随适配器改写失效：已声明 serialize_response 却仍走兜底")
+	}
+	body, err := after.NonStream(&types.UnifiedResponse{Content: "x"}, "m", "id", 1)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(body), "rewritten") {
+		t.Errorf("改写后的序列化器未生效: %s", body)
+	}
+}
+
+// TestSerializerCacheIsPerVM 是那个 bug 的回归测试：缓存曾经只按适配器名做键，
+// 于是同一进程内两个 VM 共享同一条判定——一个适配器被重写后，另一个 VM 的
+// 缓存会说「没声明钩子」，新钩子永远不生效。
+func TestSerializerCacheIsPerVM(t *testing.T) {
+	vm1 := luaVMFor(t)
+	vm2 := luaVMFor(t)
+	// 两个 VM 各自独立：改写 vm1 不得影响 vm2 的判定
+	if err := vm1.LoadAdapterSource("openai", `local a={} a.name="openai" a.version="9"
+a.endpoint="/x" function a.serialize_response(p) return '{"only_vm1":true}' end return a`); err != nil {
+		t.Fatal(err)
+	}
+	if _, f := serializerForCached(vm1, "openai").(openAISerializer); f {
+		t.Error("vm1 改写后应改走适配器序列化")
+	}
+	if _, f := serializerForCached(vm2, "openai").(openAISerializer); !f {
+		t.Error("vm2 未被改写，却因共享缓存而丢失了兜底判定（跨 VM 串扰）")
+	}
+}
+
+// ---- 并发安全 ----
+
+// TestSerializerForConcurrent 是压力测试：多 goroutine 并发取序列化器
+// （缓存是 sync.Map，HasFn 会编译脚本 —— 这里验证并发下不 panic、不返回
+// 半初始化对象，且每次结果自洽）。
+func TestSerializerForConcurrent(t *testing.T) {
+	vm := luaVMFor(t)
+	const workers, iters = 16, 40
+	var wg sync.WaitGroup
+	errs := make(chan string, workers*iters)
+	for w := 0; w < workers; w++ {
+		wg.Add(1)
+		go func(w int) {
+			defer wg.Done()
+			for i := 0; i < iters; i++ {
+				name := "openai"
+				if w%2 == 1 {
+					name = "anthropic"
+				}
+				ser := serializerForCached(vm, name)
+				if ser == nil {
+					errs <- "nil serializer"
+					return
+				}
+				// 结果必须自洽：同一适配器每次拿到同一类型，且形状与该适配器
+				// 声明的协议一致（openai→OpenAI，anthropic→Anthropic）。
+				_, isFallback := ser.(openAISerializer)
+				wantShape := "chat.completion"
+				switch name {
+				case "anthropic":
+					if isFallback {
+						errs <- "anthropic 并发下退回兜底"
+						return
+					}
+					wantShape = `"type":"message"`
+				default:
+					if !isFallback {
+						errs <- name + " 并发下未走兜底"
+						return
+					}
+				}
+				body, err := ser.NonStream(&types.UnifiedResponse{Content: "x"}, "m", "id", 1)
+				if err != nil {
+					errs <- "serialize: " + err.Error()
+					return
+				}
+				if !strings.Contains(string(body), wantShape) {
+					errs <- "并发下输出形状不一致: " + name + " -> " + string(body)
+					return
+				}
+			}
+		}(w)
+	}
+	wg.Wait()
+	close(errs)
+	for e := range errs {
+		t.Error(e)
+	}
+}

internal/gateway/server.go

@@ -18,8 +18,8 @@ import (
 	"sync"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 //go:embed ui/*
@@ -187,6 +187,10 @@ type statusRecorder struct {
 	code int
 }
 
+// Unwrap preserves optional writer capabilities (notably Hijack for WebSocket)
+// through the access-log middleware via http.ResponseController.
+func (s *statusRecorder) Unwrap() http.ResponseWriter { return s.ResponseWriter }
+
 func (s *statusRecorder) WriteHeader(code int) {
 	if s.code == 0 {
 		s.code = code
@@ -216,6 +220,14 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 	switch {
 	case r.URL.Path == "/v1/chat/completions":
 		g.handleChat(w, r)
+	case r.URL.Path == "/v1/messages":
+		// Anthropic Messages protocol. Scheduling, quotas, cooldowns, AUTO
+		// chain and accounting are EXACTLY the ones /v1/chat/completions uses —
+		// only the wire shapes differ, and those come from the serving
+		// adapter's serialize_* hooks. The kernel holds no Anthropic field
+		// names (see serializer.go: the fallback is OpenAI, and any adapter
+		// that declares serialize_* replaces it wholesale).
+		g.handleMessages(w, r)
 	case r.URL.Path == "/v1/images/generations":
 		g.handleImage(w, r)
 	case r.URL.Path == "/v1/models":
@@ -224,21 +236,27 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 		g.handleAdaptersAPI(w, r)
 	case r.URL.Path == "/api/v1" || strings.HasPrefix(r.URL.Path, "/api/v1/"):
 		g.apiV1Routes(w, r)
+	// Must precede the /api/sources prefix match: a generic HasPrefix on
+	// "/api/sources/" would otherwise swallow "/api/sources/discover" and the
+	// operator would get "source name required in path" instead of a catalog.
+	case r.URL.Path == "/api/sources/discover" || strings.HasPrefix(r.URL.Path, "/api/sources/discover/"):
+		g.handleSourceDiscoverAPI(w, r)
 	case r.URL.Path == "/api/sources" || strings.HasPrefix(r.URL.Path, "/api/sources/"):
 		g.handleSourcesAPI(w, r)
 	case r.URL.Path == "/api/source_templates" || strings.HasPrefix(r.URL.Path, "/api/source_templates/"):
 		g.handleSourceTemplatesAPI(w, r)
 	case r.URL.Path == "/api/ui-inject" || strings.HasPrefix(r.URL.Path, "/api/ui-inject/"):
 		g.handlePluginUI(w, r)
-	case r.URL.Path == "/api/plugins/billing/rules":
-		// Checked BEFORE the generic /api/plugins/ prefix: the plugin router
-		// treats anything after the name as a plugin sub-resource and would
-		// 404 on "rules" instead of reaching the billing editor.
-		g.handleBillingRules(w, r)
 	case strings.HasPrefix(r.URL.Path, pluginHTTPPrefix):
 		// Before the generic plugin router: these paths belong to a PLUGIN, not to
 		// the plugin manager, and "http" is a valid sub-resource name otherwise.
 		g.handlePluginHTTP(w, r)
+	case strings.HasPrefix(r.URL.Path, "/plugins/") || strings.HasPrefix(r.URL.Path, "/v1/plugins/"):
+		// Custom mounts a plugin or adapter claimed (R3). Checked AFTER every
+		// kernel route so a claim can never shadow one — ValidateCustomMount
+		// refuses the overlap at load time, and this ordering is the second line
+		// of defence for a table that somehow slipped through.
+		g.handleCustomMount(w, r)
 	case r.URL.Path == "/api/plugins" || strings.HasPrefix(r.URL.Path, "/api/plugins/"):
 		g.handlePluginsAPI(w, r)
 	case r.URL.Path == "/api/chat":

internal/gateway/source_fields_test.go

@@ -9,8 +9,8 @@ import (
 	"testing"
 	"time"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
-	"llmsproxy/internal/core"
+	"modelrouter/internal/core"
 )
 
 // This file covers the second round of the same class of defect bf0657b fixed.

internal/gateway/sources_api_test.go

@@ -9,7 +9,7 @@ import (
 	"strings"
 	"testing"
 
-	"llmsproxy/internal/config"
+	"modelrouter/internal/config"
 )
 
 // These cover the two write-path defects found while adding a model to a live
@@ -71,11 +71,6 @@ func TestSourcePutUpdatesSource(t *testing.T) {
 			continue
 		}
 		found = true
-		for _, m := range s.Models {
-			if m.ID == "m-one" && m.Priority != 42 {
-				t.Fatalf("priority not applied: %+v", s.Models)
-			}
-		}
 	}
 	if !found {
 		t.Fatal("source vanished after PUT")