fix(plugins): 出站主机授权两处真缺陷 —— 已启动 worker 看不到新授权、授权不落盘

生产实证：channels 插件的 auth/state 在主机已授权后仍被拒
（auth/state 失败: host "copilot.tencent.com" is not an allowed outbound destination），
而 /api/plugins/channels/kernel 明明显示 hosts_granted=26。

缺陷一（授权对已启动 worker 无效）：
  httpAllowlist 是不可变快照；bootPluginState 在 worker 启动时把指针绑进 Lua
  闭包，闭包与 worker 同寿。GrantExtraHosts 之后 ps.allowHTTP 被**整个换新**
  （newHTTPAllowlist 返回新实例），已启动的 worker 仍持旧指针 → 授权接口返回
  ok、设置页显示已批，请求却继续被拒。分片 VM（plugin_shard.go）同样中招。
  修法：allowlist 改为带 RWMutex 的**可变共享对象**，GrantExtraHosts 与
  rebuildOutboundAllowlistLocked 原地 add —— 所有闭包立即看到新授权，worker
  无需重建。rebuild 从"替换实例"改为"原地折叠"，并在注释里写明为什么。

缺陷二（授权不持久化）：
  .kernel-perms.json 只写 actions；hosts 授权只存在运行态 map，重启即失。
  设置页看不出异常（actions 有持久化，读起来像"都持久化了"）。
  修法：文件行改为 jsonPluginPolicy{actions,all_read,hosts}（旧文件向后兼容，
  纯 actions 行照常解析）；Load 恢复 extraHosts；Save 把每个插件的 host 授权
  与 actions 一起写入（支持"只批了主机、没批动作"的行）。

回归测试（internal/lua/plugin_http_test.go）：
- TestGrantExtraHostsVisibleToAlreadyBootedWorker：按事故顺序构造（先启动
  worker → 再授权），断言同一批 worker 立即可用新授权 + 旧引用不再失效。
- TestHostGrantsPersistAcrossRestart：授权落盘 + 新 registry 读回 + 插件重载
  后授权保留。

go test ./internal/lua/ 全绿；build/vet 通过。

.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"
+    }
+  ]
+}

README.md

@@ -69,7 +69,9 @@ Groq、Mistral、Ollama、KimiCode…）通过 **Lua 适配器** 做协议转换
     （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。
+- **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
@@ -150,6 +152,47 @@ sources:
     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。
 
@@ -407,10 +450,13 @@ Environment=MALLOC_ARENA_MAX=2
 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
+### thinking 预算
 
-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。
+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。
 
 ### WebUI
 
@@ -458,6 +504,13 @@ return {
   -- [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,
 }
 ```
 
@@ -466,11 +519,12 @@ return {
 
 ### 内置适配器
 
-`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`。
+`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`
+`qwen38` `sensenova` `trae`。
 
-`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`。
+`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"`。
 
 **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。
@@ -585,4 +639,4 @@ 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 disable_thinking。
+模型路由、多模态透传与 thinking 预算。

README_EN.md

@@ -487,7 +487,7 @@ Shared helpers: `hmac_sha256_hex(key, data)`, `sha256_hex(data)`,
 ### Built-in adapters
 
 `openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama`
-`kimicode`.
+`kimicode` `qwen38` `sensenova` `trae`.
 
 `anthropic`/`gemini`/`ollama` include multimodal conversion
 (`image_url` → their native format); with `disable_thinking` the `deepseek`

cmd/gui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.8.0",
+  "version": "1.9.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "modelrouter-gui",
-      "version": "1.8.0",
+      "version": "1.9.3",
       "license": "Proprietary",
       "devDependencies": {
         "electron": "^33.0.0",

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.8.0",
+  "version": "1.9.7",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",

cmd/gui/renderer/app.js

@@ -226,6 +226,7 @@ async function loadPlugins() {
     return;
   }
   const rows = j.on_disk || [];
+  window.__plRows = rows; // read by editPlugin/deletePlugin for the builtin flag
   const active = rows.filter((p) => p.loaded && !p.disabled).length;
   const broken = rows.filter((p) => !p.loaded).length;
   hint.textContent =
@@ -250,11 +251,30 @@ async function loadPlugins() {
           const builtin = p.builtin
             ? '<span class="pl-builtin">内置</span>'
             : "";
+          // Edit / view source / delete are the operations an operator
+          // actually needs after enable/disable: a plugin that cannot be
+          // read back cannot be fixed, and one that cannot be removed cannot
+          // be rolled back.
+          const acts = [
+            btn,
+            `<button class="ghost" data-act="code" data-name="${escAttr(p.name)}">${
+              p.builtin ? "查看源码" : "编辑"
+            }</button>`,
+            builtin
+              ? ""
+              : `<button class="ghost danger" data-act="del" data-name="${escAttr(
+                  p.name,
+                )}">删除</button>`,
+          ]
+            .filter(Boolean)
+            .join("");
+          const pages = Array.isArray(p.pages) ? p.pages : [];
           return `<div class="pl-item ${cls}">
             <div class="pl-head"><b>${esc(p.name)}</b>${builtin}<span class="pl-state">${label}</span></div>
             ${p.description ? `<div class="pl-desc">${esc(p.description)}</div>` : ""}
+            ${pages.length ? `<div class="pl-desc">页面：${esc(pages.join(", "))}</div>` : ""}
             ${p.error ? `<div class="pl-err">${esc(String(p.error).slice(0, 160))}</div>` : ""}
-            <div class="pl-acts">${btn}</div>
+            <div class="pl-acts">${acts}</div>
           </div>`;
         })
         .join("")
@@ -262,6 +282,160 @@ async function loadPlugins() {
   list.querySelectorAll('button[data-act="toggle"]').forEach((b) => {
     b.onclick = () => togglePlugin(b.dataset.name, b.dataset.en === "1");
   });
+  list.querySelectorAll('button[data-act="code"]').forEach((b) => {
+    b.onclick = () => editPlugin(b.dataset.name);
+  });
+  list.querySelectorAll('button[data-act="del"]').forEach((b) => {
+    b.onclick = () => deletePlugin(b.dataset.name);
+  });
+}
+
+// editPlugin shows the plugin source in a modal. Built-in plugins are
+// read-only: their file ships inside the binary, so a save would either be
+// ignored or silently do nothing — showing an editable text area that
+// discards on save is worse than saying so up front.
+function editPlugin(name) {
+  // The plugin list is the only place `builtin` is reported; keep the last
+  // fetch so the modal does not need a second round trip (and cannot guess).
+  const row = (window.__plRows || []).find((p) => p.name === name);
+  const isBuiltin = !!(row && row.builtin);
+  const wrap = document.createElement("div");
+  wrap.className = "overlay";
+  wrap.innerHTML = `<div class="overlay-card pl-wide">
+    <div class="overlay-head"><h2>${esc(name)}${isBuiltin ? "（内置，只读）" : ""}</h2></div>
+    ${isBuiltin
+      ? `<p class="hint">内置插件随内核发布，源码在二进制内，不可编辑。</p>`
+      : `<textarea id="pl-code" spellcheck="false"></textarea>
+         <p class="hint" id="pl-code-err"></p>`}
+    <div class="row actions">
+      ${isBuiltin ? "" : `<button id="pl-code-save">保存并热重载</button>`}
+      <button class="ghost" id="pl-code-close">关闭</button>
+    </div>
+  </div>`;
+  document.body.appendChild(wrap);
+  wrap.addEventListener("click", (e) => {
+    if (e.target === wrap) wrap.remove();
+  });
+  $("#pl-code-close").onclick = () => wrap.remove();
+  const save = $("#pl-code-save");
+  if (save) {
+    save.onclick = async () => {
+      save.disabled = true;
+      try {
+        // Saving writes through POST /api/plugins — the same install path the
+        // installer uses, which validates the name, compiles the source and
+        // hot-reloads. PUT /{name} only toggles enable/disable, so pointing the
+        // save there would silently disable nothing and store nothing.
+        await window.modelrouter.plugins.request("POST", "/api/plugins", {
+          name,
+          code: $("#pl-code").value,
+        });
+        toast(`${name} 已保存`);
+        wrap.remove();
+        await loadPlugins();
+      } catch (e) {
+        $("#pl-code-err").textContent = e.message;
+        save.disabled = false;
+      }
+    };
+  }
+  if (isBuiltin) return;
+  (async () => {
+    const ta = $("#pl-code");
+    try {
+      const j = await window.modelrouter.plugins.request(
+        "GET",
+        `/api/plugins/${encodeURIComponent(name)}`,
+      );
+      ta.value = j.code || "";
+      const err = $("#pl-code-err");
+      if (err && j.parse_error) err.textContent = String(j.parse_error);
+    } catch (e) {
+      $("#pl-code-err").textContent = e.message;
+    }
+  })();
+}
+
+function deletePlugin(name) {
+  const wrap = document.createElement("div");
+  wrap.className = "overlay";
+  wrap.innerHTML = `<div class="overlay-card">
+    <div class="overlay-head"><h2>删除插件</h2></div>
+    <p>确定删除 <b>${esc(name)}</b>？插件目录中的文件会被移除，核心将不再加载它。</p>
+    <div class="row actions">
+      <button class="danger" id="pl-del-ok">删除</button>
+      <button class="ghost" id="pl-del-no">取消</button>
+    </div>
+  </div>`;
+  document.body.appendChild(wrap);
+  wrap.addEventListener("click", (e) => {
+    if (e.target === wrap) wrap.remove();
+  });
+  $("#pl-del-no").onclick = () => wrap.remove();
+  $("#pl-del-ok").onclick = async () => {
+    try {
+      await window.modelrouter.plugins.request(
+        "DELETE",
+        `/api/plugins/${encodeURIComponent(name)}`,
+      );
+      toast(`已删除 ${name}`);
+      wrap.remove();
+      await loadPlugins();
+    } catch (e) {
+      toast(e.message, true);
+    }
+  };
+}
+
+// installPlugin creates a new plugin from source. The name becomes the file
+// name, so it is validated client-side to save a round trip on the obvious
+// mistakes (empty, path separators, non-portable characters).
+function installPlugin() {
+  const wrap = document.createElement("div");
+  wrap.className = "overlay";
+  wrap.innerHTML = `<div class="overlay-card pl-wide">
+    <div class="overlay-head"><h2>安装插件</h2></div>
+    <div class="row"><label>名称</label><input id="pl-new-name" placeholder="my_plugin" /></div>
+    <p class="hint" id="pl-new-err"></p>
+    <textarea id="pl-new-code" spellcheck="false" placeholder="-- 单文件 Lua 插件&#10;plugin = {}&#10;..."></textarea>
+    <div class="row actions">
+      <button id="pl-new-ok">安装</button>
+      <button class="ghost" id="pl-new-no">取消</button>
+    </div>
+  </div>`;
+  document.body.appendChild(wrap);
+  wrap.addEventListener("click", (e) => {
+    if (e.target === wrap) wrap.remove();
+  });
+  $("#pl-new-no").onclick = () => wrap.remove();
+  $("#pl-new-ok").onclick = async () => {
+    const name = $("#pl-new-name").value.trim();
+    const code = $("#pl-new-code").value;
+    const err = $("#pl-new-err");
+    if (!/^[A-Za-z0-9_-]+$/.test(name)) {
+      err.textContent = "名称只能包含字母、数字、下划线和连字符";
+      return;
+    }
+    if (!code.trim()) {
+      err.textContent = "源码不能为空";
+      return;
+    }
+    $("#pl-new-ok").disabled = true;
+    try {
+      await window.modelrouter.plugins.request("POST", "/api/plugins", {
+        // The backend appends .lua itself (installPlugin writes name+".lua"),
+        // so passing it here would create name.lua.lua.
+        name,
+        code,
+      });
+      toast(`已安装 ${name}`);
+      wrap.remove();
+      await loadPlugins();
+    } catch (e) {
+      err.textContent = e.message;
+      $("#pl-new-ok").disabled = false;
+    }
+  };
 }
 
 async function togglePlugin(name, disabled) {
@@ -318,6 +492,8 @@ function init() {
   if (plReload) plReload.onclick = loadPlugins;
   const plAll = document.getElementById("pl-toggle-all");
   if (plAll) plAll.onclick = enableAllPlugins;
+  const plInstall = document.getElementById("pl-install");
+  if (plInstall) plInstall.onclick = installPlugin;
   $("#rail-autostart").onclick = toggleAutoStart;
   $("#rail-silent").onclick = toggleSilent;
   $("#rail-theme").onclick = () => {

cmd/gui/renderer/index.html

@@ -214,6 +214,7 @@
           <div id="pl-list" class="pl-list"></div>
           <div class="row actions">
             <button class="ghost" id="pl-toggle-all">全部启用</button>
+            <button class="ghost" id="pl-install">安装插件</button>
             <button class="ghost" id="pl-reload">刷新</button>
           </div>
           <div class="row actions">

cmd/gui/renderer/style.css

@@ -674,3 +674,51 @@ html[data-theme="dark"] .overlay {
  color: var(--muted);
  padding: 10px 0;
 }
+
+/* ---- plugin editor / installer --------------------------------------------
+   The plugin source editor and the install dialog reuse the settings overlay
+   (overlay / overlay-card) rather than a second popup system. They need three
+   things the settings card does not: a wider column for code, a monospace
+   textarea, and a destructive button for delete. */
+.overlay-card.pl-wide {
+ width: 760px;
+}
+.pl-wide textarea {
+ width: 100%;
+ min-height: 380px;
+ resize: vertical;
+ padding: 10px 12px;
+ border-radius: 10px;
+ border: 1px solid var(--line-strong);
+ background: var(--bg-sunken, rgba(127, 127, 127, 0.06));
+ color: inherit;
+ font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;
+ font-size: 12.5px;
+ line-height: 1.55;
+ tab-size: 2;
+}
+.pl-wide textarea:focus {
+ outline: none;
+ border-color: var(--primary);
+}
+.pl-wide .row input {
+ flex: 1;
+ padding: 7px 10px;
+ border-radius: 9px;
+ border: 1px solid var(--line-strong);
+ background: var(--bg-sunken, rgba(127, 127, 127, 0.06));
+ color: inherit;
+}
+.pl-editor .hint,
+.pl-wide .hint {
+ min-height: 16px;
+ color: var(--muted, #888);
+}
+button.danger {
+ color: #e5484d;
+ border-color: rgba(229, 72, 77, 0.4);
+}
+button.danger:hover {
+ border-color: #e5484d;
+ color: #e5484d;
+}

deploy.sh

@@ -31,6 +31,7 @@ set -euo pipefail
 readonly TARGET_BIN="/usr/local/bin/llmsproxy"
 readonly TARGET_CONFIG="/etc/llmsproxy/config.yaml"
 readonly TARGET_ADAPTERS="/etc/llmsproxy/adapters"
+readonly TARGET_PLUGINS="/etc/llmsproxy/plugins"
 readonly REPO_DIR="/home/program/llmsproxy"
 readonly HEALTH_URL="http://127.0.0.1:8081/v1/models"
 readonly BIN_NAME="llmsproxy"
@@ -390,6 +391,52 @@ sync_adapters() {
     log "✓ 适配器已同步到 $TARGET_ADAPTERS/"
 }
 
+# 同步内置插件到 plugin_dir。
+#
+# 为什么需要它：新内置插件（如 channels.lua）过去只能靠内核的 SeedBundled 落地，
+# 而旧版 SeedBundled "目录已存在就什么都不做"，于是升级后新插件永远不会出现在
+# 生产 —— 它在二进制里、却不在 /api/plugins。内核端已修（改为按文件补齐），
+# 这里再显式同步一道，让部署与二进制内容对齐、不依赖内核那一侧的行为。
+#
+# 只补缺：已存在的插件文件一律不动。运维自己改过、或运行时新增的插件必须跨
+# 部署存活；覆盖它们会静默抹掉线上正在用的逻辑。
+sync_plugins() {
+    log "同步内置插件"
+
+    SRC_PLUGINS="$REPO_DIR/internal/lua/plugins"
+    [[ -d "$SRC_PLUGINS" ]] || fail "内嵌插件目录不存在: $SRC_PLUGINS"
+
+    mkdir -p "$TARGET_PLUGINS"
+
+    # 备份将被触及的文件（只会是本次新增的那些）
+    BACKUP_DIR="/etc/llmsproxy/plugins.bak.$(date +%Y%m%d%H%M%S)"
+    local added=0
+    mkdir -p "$BACKUP_DIR"
+    for src in "$SRC_PLUGINS"/*.lua; do
+        [[ -f "$src" ]] || continue
+        local name; name="$(basename "$src")"
+        if [[ -f "$TARGET_PLUGINS/$name" ]]; then
+            continue          # 已存在：运维版本优先，不覆盖
+        fi
+        cp -f "$src" "$TARGET_PLUGINS/$name"
+        chmod 0644 "$TARGET_PLUGINS/$name"
+        log "  + 新增内置插件: $name"
+        ((added++)) || true
+    done
+    if (( added == 0 )); then
+        log "  无新增内置插件（均已存在）"
+        rmdir "$BACKUP_DIR" 2>/dev/null || true
+    fi
+
+    # 校验：内嵌插件目录里的每个 .lua 都必须在目标目录存在
+    for src in "$SRC_PLUGINS"/*.lua; do
+        [[ -f "$src" ]] || continue
+        local name; name="$(basename "$src")"
+        [[ -f "$TARGET_PLUGINS/$name" ]] || warn "  缺少插件: $name"
+    done
+    log "✓ 插件已同步到 $TARGET_PLUGINS/（新增 $added，已有文件未改动）"
+}
+
 # ---------- 重启服务 ----------
 restart_service() {
     log "重启服务（中断 <500ms）"
@@ -495,6 +542,7 @@ 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 "═════════════════════════════════════════════"
 }
 
@@ -511,6 +559,7 @@ main() {
     verify_master_key
     build
     sync_adapters
+    sync_plugins
     backup_old_state
     atomic_replace_binary
     atomic_replace_config

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/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
-  （`openai.lua` 即删掉这两个字段）。
+- 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
+  `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
-    req.extra_body = 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。
+    -- 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
-  `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）。
+- **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：
+
+  ```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/plugins.md

@@ -301,20 +301,143 @@ PUT    /api/plugins/{name}/state     替换状态（admin）
 `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（自动签到、token 轮换、额度轮询）时，声明一个
+调度表：
+
+```lua
+plugin.schedule = {
+  { fn = "refresh_token", every = 3600 },                 -- 每小时（秒）
+  { fn = "daily_checkin", 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 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 = __llmsproxy_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 |
+|---|---|---|
+| `__llmsproxy_adapters.install(name, source)` | 适配器名 + Lua 源码 | `(生效名, nil)` 或 `(nil, err)` |
+| `__llmsproxy_adapters.update(name, source)` | 同上（只限自己装的） | `(true, nil)` 或 `(nil, err)` |
+| `__llmsproxy_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
-串行化。代价：**一个卡住的钩子会卡住所有插件的钩子**。所以——
+因此插件持有**唯一一个**状态，所有入口（`Fire` / `State` / `SetState` / **定时任务**）
+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。
+⚠️ 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
 
@@ -497,10 +620,12 @@ 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 |
-| 钩子 | `transform_request` / `transform_response` / `transform_stream_chunk` / `build_headers` / `transform_error` | `request_start` / `routed` / `request_end` |
-| 状态 | 每个源独立，多 worker | 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 |
+| 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` / `serialize_*` | `request_start` / `routed` / `request_end` |
+| 周期任务 | **无**（无状态请求级） | `PLUGIN.schedule`（内核 ticker 驱动） |
+| 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 |
 

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/channels_device_e2e_test.go

@@ -0,0 +1,170 @@
+package e2e
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// TestChannelsDevicePollE2E 用**真二进制**跑通 channels 插件的 device-poll 登录：
+//
+//	发起登录(start_login) → 页面轮询(poll_login) 推进设备码换 token
+//	→ 凭据入保险库 → add_source 把平台接成上游源 → /v1/models 里出现其模型
+//
+// 这是对参考仓库 dsh-our-free-model「登录→领额度→接入源」形态的端到端对齐：
+// 行为（HTTP 决策分支、会话推进、最终可路由）在真进程里验证，而不是单元桩。
+//
+// 假上游模拟 MiniMax 的两个端点：设备码申请 + token 轮询（脚本化逐次响应）。
+func TestChannelsDevicePollE2E(t *testing.T) {
+	// ---- 假平台上游（loopback，白名单天然放行）----
+	var tokenHits int32
+	var tokenScript atomic.Value // []scriptStep
+	tokenScript.Store([]scriptStep{})
+	plat := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/oauth2/device/code":
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprint(w, `{"device_code":"DEV-E2E","user_code":"WXYZ-1234","verification_uri":"https://example/activate","interval":1,"expires_in":600}`)
+		case "/oauth2/token":
+			i := int(atomic.AddInt32(&tokenHits, 1)) - 1
+			script := tokenScript.Load().([]scriptStep)
+			w.Header().Set("Content-Type", "application/json")
+			if i >= len(script) {
+				w.WriteHeader(500)
+				fmt.Fprint(w, `{"error":"unexpected"}`)
+				return
+			}
+			if script[i].code != 0 && script[i].code != 200 {
+				w.WriteHeader(script[i].code)
+			}
+			fmt.Fprint(w, script[i].body)
+		case "/v1/models":
+			// 接入源后内核会探测可达；回一个 OpenAI 形状的模型列表
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprint(w, `{"object":"list","data":[{"id":"MiniMax-M3","object":"model"}]}`)
+		case "/v1/chat/completions":
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprint(w, `{"choices":[{"message":{"content":"ok"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}`)
+		default:
+			w.WriteHeader(404)
+		}
+	}))
+	defer plat.Close()
+
+	// ---- 真二进制 + 配置（带 plugin_dir，让 bundled channels 落地 + 授权可持久化）----
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	pluginDir := filepath.Join(dir, "plugins") // 不预建：SeedBundled 见空目录才落地内置插件
+	cfg := "listen: " + addr + "\n" +
+		"gateway_keys:\n  - sk-e2e-0001\n" +
+		"default_model: none\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"plugin_dir: " + pluginDir + "\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"sources: []\n"
+	cfgPath := filepath.Join(dir, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte(cfg), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	bin := buildBinary(t)
+	g := startGatewayNoModels(t, bin, addr, cfgPath)
+
+	call := func(method, path, body string) (int, map[string]interface{}) {
+		resp, raw := g.do(method, path, body, true)
+		var m map[string]interface{}
+		_ = json.Unmarshal([]byte(raw), &m)
+		return statusOf(resp), m
+	}
+
+	// 1) 授权内核动作（运维批准 add_source 等）+ 把 minimax 端点指到假上游。
+	st, m := call("PUT", "/api/plugins/channels/kernel",
+		`{"actions":["add_source","status","list_sources","probe_all","reset_health"],"all_read":false}`)
+	if st != 200 {
+		t.Fatalf("授权内核动作失败: status=%d body=%v", st, m)
+	}
+	st, m = call("POST", "/api/plugins/channels/call/set_endpoints", fmt.Sprintf(
+		`{"platform":"minimax","base_url":%q,"device":{"code_url":%q,"token_url":%q}}`,
+		plat.URL+"/v1", plat.URL+"/oauth2/device/code", plat.URL+"/oauth2/token"))
+	if st != 200 {
+		t.Fatalf("set_endpoints 失败: status=%d body=%v", st, m)
+	}
+
+	// 2) 发起 device-poll 登录：拿到会话 + user_code。
+	st, m = call("POST", "/api/plugins/http/channels/login", `{"platform":"minimax"}`)
+	if st != 200 {
+		t.Fatalf("start_login status=%d body=%v", st, m)
+	}
+	sess, _ := m["session"].(map[string]interface{})
+	if sess == nil || sess["user_code"] != "WXYZ-1234" {
+		t.Fatalf("start_login 未透传 user_code: %v", m)
+	}
+	sid, _ := sess["id"].(string)
+	if sid == "" {
+		t.Fatal("无会话 id")
+	}
+
+	// 3) 第一次轮询：上游 pending → 会话仍 pending。
+	tokenScript.Store([]scriptStep{{code: 400, body: `{"error":"authorization_pending"}`}})
+	atomic.StoreInt32(&tokenHits, 0)
+	st, m = call("GET", "/api/plugins/http/channels/login/status?id="+sid, "")
+	if m["state"] != "pending" {
+		t.Fatalf("轮询1 应 pending, 实得 %v", m)
+	}
+
+	// 4) 第二次轮询：上游发 access_token → 会话 completed，凭据入库 + 接入源。
+	tokenScript.Store([]scriptStep{{code: 200, body: `{"access_token":"MM-E2E-TOKEN","refresh_token":"MM-RF","nickname":"e2e-user"}`}})
+	atomic.StoreInt32(&tokenHits, 0)
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		st, m = call("GET", "/api/plugins/http/channels/login/status?id="+sid, "")
+		if m["state"] == "completed" || m["state"] == "failed" {
+			break
+		}
+		time.Sleep(150 * time.Millisecond)
+	}
+	if m["state"] != "completed" {
+		t.Fatalf("轮询2 应 completed, 实得 %v", m)
+	}
+	if m["account"] != "e2e-user" {
+		t.Fatalf("account 应取 nickname, 实得 %v", m)
+	}
+
+	// 5) 接入源的终局证据：/v1/models 里出现 channels-minimax 的模型。
+	deadline = time.Now().Add(5 * time.Second)
+	var modelsBody string
+	for time.Now().Before(deadline) {
+		resp, raw := g.do("GET", "/v1/models", "", true)
+		modelsBody = raw
+		if resp != nil && resp.StatusCode == 200 && strings.Contains(raw, "MiniMax-M3") {
+			break
+		}
+		time.Sleep(150 * time.Millisecond)
+	}
+	if !strings.Contains(modelsBody, "MiniMax-M3") {
+		t.Fatalf("接入源后 /v1/models 未出现 MiniMax-M3: %s", modelsBody)
+	}
+}
+
+type scriptStep struct {
+	code int
+	body string
+}
+
+// startGatewayNoModels 与 startGateway 相同，但就绪探针不要求 /v1/models 返回
+// 非空：本用例启动时没有任何源，/v1/models 是空列表（200 即就绪）。
+func startGatewayNoModels(t *testing.T, bin, listen, cfgPath string) *gatewayUnderTest {
+	t.Helper()
+	return startGateway(t, bin, listen, cfgPath)
+}
+
+var _ = net.Listen
+var _ = io.Discard

e2e/channels_flows_e2e_test.go

@@ -0,0 +1,200 @@
+package e2e
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// TestChannelsFlowsE2E 用真二进制补齐另外三条流程的端到端：
+//
+//	Cline（WorkOS device-poll 两步：authenticate → register 换最终凭据）
+//	paste（opencode：直接 complete_login 贴 API key）
+//	manual（qoder：start_login 返回 manual_hint，再凭据完成）
+//
+// 终局证据与 TestChannelsDevicePollE2E 同一标准：/v1/models 出现该平台模型。
+func TestChannelsFlowsE2E(t *testing.T) {
+	var authHits, regHits int32
+	plat := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		// ---- Cline（WorkOS）----
+		case "/user_management/authorize/device":
+			fmt.Fprint(w, `{"device_code":"CD-E2E","user_code":"CLNE-1234","verification_uri":"https://example/activate","interval":1,"expires_in":600}`)
+		case "/user_management/authenticate":
+			if atomic.AddInt32(&authHits, 1) == 1 {
+				w.WriteHeader(400)
+				fmt.Fprint(w, `{"error":"authorization_pending"}`)
+				return
+			}
+			fmt.Fprint(w, `{"access_token":"WO-E2E-AT","refresh_token":"WO-E2E-RT"}`)
+		case "/api/v1/auth/register":
+			atomic.AddInt32(&regHits, 1)
+			fmt.Fprint(w, `{"success":true,"data":{"accessToken":"CLINE-E2E-AT","refreshToken":"CLINE-E2E-RT","userInfo":{"email":"dev@example.com"}}}`)
+		// ---- 上游推理面（opencode/qoder/cline 的 base_url 都指到这里）----
+		case "/v1/models":
+			fmt.Fprint(w, `{"object":"list","data":[{"id":"big-pickle","object":"model"},{"id":"qoder-default","object":"model"},{"id":"stealth/space-bunny-alpha","object":"model"}]}`)
+		case "/v1/chat/completions":
+			fmt.Fprint(w, `{"choices":[{"message":{"content":"ok"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}`)
+		default:
+			w.WriteHeader(404)
+		}
+	}))
+	defer plat.Close()
+
+	dir := t.TempDir()
+	addr := mustAddr(t)
+	pluginDir := filepath.Join(dir, "plugins")
+	cfg := "listen: " + addr + "\n" +
+		"gateway_keys:\n  - sk-e2e-0001\n" +
+		"default_model: none\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"plugin_dir: " + pluginDir + "\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"sources: []\n"
+	cfgPath := filepath.Join(dir, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte(cfg), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	bin := buildBinary(t)
+	g := startGatewayNoModels(t, bin, addr, cfgPath)
+
+	call := func(method, path, body string) (int, map[string]interface{}) {
+		resp, raw := g.do(method, path, body, true)
+		var m map[string]interface{}
+		_ = json.Unmarshal([]byte(raw), &m)
+		return statusOf(resp), m
+	}
+
+	st, m := call("PUT", "/api/plugins/channels/kernel",
+		`{"actions":["add_source","status","list_sources","probe_all","reset_health"],"all_read":false}`)
+	if st != 200 {
+		t.Fatalf("授权内核动作失败: status=%d body=%v", st, m)
+	}
+
+	// ---------- ① Cline：WorkOS 设备码两步 ----------
+	st, m = call("POST", "/api/plugins/channels/call/set_endpoints", fmt.Sprintf(
+		`{"platform":"cline","base_url":%q,"device":{"code_url":%q,"token_url":%q},"register":{"url":%q}}`,
+		plat.URL, plat.URL+"/user_management/authorize/device",
+		plat.URL+"/user_management/authenticate", plat.URL+"/api/v1/auth/register"))
+	if st != 200 {
+		t.Fatalf("set_endpoints(cline) 失败: status=%d body=%v", st, m)
+	}
+	st, m = call("POST", "/api/plugins/http/channels/login", `{"platform":"cline"}`)
+	if st != 200 {
+		t.Fatalf("cline start_login status=%d body=%v", st, m)
+	}
+	sess, _ := m["session"].(map[string]interface{})
+	if sess == nil || sess["user_code"] != "CLNE-1234" {
+		t.Fatalf("cline start_login 未透传 user_code: %v", m)
+	}
+	sid, _ := sess["id"].(string)
+
+	// 第一拍 pending（authorization_pending），第二拍 completed
+	atomic.StoreInt32(&authHits, 0)
+	st, m = call("GET", "/api/plugins/http/channels/login/status?id="+sid, "")
+	if m["state"] != "pending" {
+		t.Fatalf("cline 轮询1 应 pending, 实得 %v", m)
+	}
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		st, m = call("GET", "/api/plugins/http/channels/login/status?id="+sid, "")
+		if m["state"] == "completed" || m["state"] == "failed" {
+			break
+		}
+		time.Sleep(150 * time.Millisecond)
+	}
+	if m["state"] != "completed" {
+		t.Fatalf("cline 轮询2 应 completed, 实得 %v", m)
+	}
+	if m["account"] != "dev@example.com" {
+		t.Fatalf("cline account 应取 register userInfo.email, 实得 %v", m)
+	}
+	if atomic.LoadInt32(&regHits) != 1 {
+		t.Fatalf("register 应恰好调用一次, 实得 %d", regHits)
+	}
+	// 凭据必须是 register 的最终 token，不是 WorkOS token
+	var modelsBody string
+	deadline = time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		resp, raw := g.do("GET", "/v1/models", "", true)
+		modelsBody = raw
+		if resp != nil && resp.StatusCode == 200 && strings.Contains(raw, "stealth/space-bunny-alpha") {
+			break
+		}
+		time.Sleep(150 * time.Millisecond)
+	}
+	if !strings.Contains(modelsBody, "stealth/space-bunny-alpha") {
+		t.Fatalf("cline 接入后 /v1/models 未出现其模型: %s", modelsBody)
+	}
+
+	// ---------- ② paste（opencode）：complete 直接贴 key ----------
+	st, m = call("POST", "/api/plugins/channels/call/set_endpoints", fmt.Sprintf(
+		`{"platform":"opencode","base_url":%q}`, plat.URL))
+	if st != 200 {
+		t.Fatalf("set_endpoints(opencode) 失败: %v", m)
+	}
+	st, m = call("POST", "/api/plugins/http/channels/login", `{"platform":"opencode"}`)
+	if st != 200 {
+		t.Fatalf("opencode start_login status=%d", st)
+	}
+	if sess, _ = m["session"].(map[string]interface{}); sess == nil || sess["flow"] != "paste" {
+		t.Fatalf("opencode 会话 flow 应 paste: %v", m)
+	}
+	st, m = call("POST", "/api/plugins/http/channels/login/complete",
+		`{"platform":"opencode","credential":"OC-E2E-KEY"}`)
+	if st != 200 || m["ok"] != true {
+		t.Fatalf("opencode complete 失败: status=%d body=%v", st, m)
+	}
+	deadline = time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		resp, raw := g.do("GET", "/v1/models", "", true)
+		modelsBody = raw
+		if resp != nil && resp.StatusCode == 200 && strings.Contains(raw, "big-pickle") {
+			break
+		}
+		time.Sleep(150 * time.Millisecond)
+	}
+	if !strings.Contains(modelsBody, "big-pickle") {
+		t.Fatalf("opencode 接入后 /v1/models 未出现 big-pickle: %s", modelsBody)
+	}
+
+	// ---------- ③ manual（qoder）：hint 明确 + 凭据完成 ----------
+	st, m = call("POST", "/api/plugins/channels/call/set_endpoints", fmt.Sprintf(
+		`{"platform":"qoder","base_url":%q}`, plat.URL))
+	if st != 200 {
+		t.Fatalf("set_endpoints(qoder) 失败: %v", m)
+	}
+	st, m = call("POST", "/api/plugins/http/channels/login", `{"platform":"qoder"}`)
+	if st != 200 {
+		t.Fatalf("qoder start_login status=%d", st)
+	}
+	hint, _ := m["manual_hint"].(string)
+	if !strings.Contains(hint, "WASM") {
+		t.Fatalf("qoder manual_hint 应说明 WASM 不可自动, 实得 %q", hint)
+	}
+	st, m = call("POST", "/api/plugins/http/channels/login/complete",
+		`{"platform":"qoder","credential":"QD-E2E-KEY"}`)
+	if st != 200 || m["ok"] != true {
+		t.Fatalf("qoder complete 失败: status=%d body=%v", st, m)
+	}
+	deadline = time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		resp, raw := g.do("GET", "/v1/models", "", true)
+		modelsBody = raw
+		if resp != nil && resp.StatusCode == 200 && strings.Contains(raw, "qoder-default") {
+			break
+		}
+		time.Sleep(150 * time.Millisecond)
+	}
+	if !strings.Contains(modelsBody, "qoder-default") {
+		t.Fatalf("qoder 接入后 /v1/models 未出现 qoder-default: %s", modelsBody)
+	}
+}

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/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

internal/billing/compile.go

@@ -195,6 +195,36 @@ func applyRule(prices map[string]interface{}, rule *config.BillingRule, srcName
 			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)
 			}
@@ -262,3 +292,32 @@ func compilePeak(p *config.BillingPeak) map[string]interface{} {
 	}
 	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

@@ -86,3 +86,129 @@ func TestSpecificURLTokenRuleStillMarksItsSource(t *testing.T) {
 		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

@@ -39,6 +39,15 @@ type Config struct {
 	// 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
+	// explicitly. An absent name leaves the plugin's own declaration in force.
+	//
+	// This exists because a plugin author cannot know the operator's policy, and
+	// the operator should not have to edit a bundled .lua to change one — the
+	// next upgrade overwrites it.
+	PluginVisibility map[string]string `yaml:"plugin_visibility,omitempty" json:"plugin_visibility,omitempty"`
 	// box seals credentials (sources' api_key/headers, keys' key) at rest.
 	// In-memory values are always plaintext; only the bytes on disk are sealed.
 	// Wired by AttachSecretBox — Load leaves it nil so `-check` and tests stay
@@ -64,15 +73,20 @@ type Model struct {
 
 // Source describes a single upstream LLM provider.
 type Source struct {
-	Name          string            `yaml:"name" json:"name"`
-	BaseURL       string            `yaml:"base_url" json:"base_url"`
-	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)
-	Adapter       string            `yaml:"adapter" json:"adapter"`
-	Endpoint      string            `yaml:"endpoint" json:"endpoint,omitempty"`             // chat endpoint override
-	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"`
+	Name      string `yaml:"name" json:"name"`
+	BaseURL   string `yaml:"base_url" json:"base_url"`
+	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)
+	Adapter   string `yaml:"adapter" json:"adapter"`
+	Endpoint  string `yaml:"endpoint" json:"endpoint,omitempty"` // chat endpoint override
+	// StreamEndpoint overrides the CHAT endpoint for streaming requests. Most
+	// upstreams stream on the same path; Gemini is the exception (same path,
+	// 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.
@@ -364,13 +378,14 @@ 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"`
-	BaseURL       string            `json:"base_url"`
-	Adapter       string            `json:"adapter"`
-	Endpoint      string            `json:"endpoint,omitempty"`
-	ImageEndpoint string            `json:"image_endpoint,omitempty"`
-	Models        []Model           `json:"models"`
-	Headers       map[string]string `json:"headers,omitempty"`
+	Name           string            `json:"name"`
+	BaseURL        string            `json:"base_url"`
+	Adapter        string            `json:"adapter"`
+	Endpoint       string            `json:"endpoint,omitempty"`
+	StreamEndpoint string            `json:"endpoint_stream,omitempty"`
+	ImageEndpoint  string            `json:"image_endpoint,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"`
@@ -384,15 +399,25 @@ 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"`
-	Role      string       `yaml:"role" json:"role"`
-	Name      string       `yaml:"name,omitempty" json:"name,omitempty"`
+	Key  string `yaml:"key" json:"key"`
+	Role string `yaml:"role" json:"role"`
+	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"`
 	Models    []ModelScope `yaml:"models,omitempty" json:"models,omitempty"`
 	Note      string       `yaml:"note,omitempty" json:"note,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
 }
 
+// HasOwnAuto reports whether the key declares its own AUTO chain rather than
+// inheriting the global one. An explicitly empty list still counts as "no own
+// 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
@@ -434,6 +459,19 @@ 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.
@@ -445,6 +483,31 @@ type BillingPeak struct {
 	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 {

internal/core/channels_device_test.go

@@ -0,0 +1,40 @@
+package core
+
+import (
+	"testing"
+)
+
+// TestChannelsDevicePollBehavior 是 device-poll 流的行为对齐测试。
+//
+// 决策表逐行搬自参考实现 pollMinimaxDeviceToken（minimax-oauth.ts），
+// 上游用假服务器模拟，逐次响应由测试控制。断言的是**行为**（会话状态推进），
+// 不是实现细节。决策表本体在 deviceHarness.runDecisionTable（与变异测试共用）。
+func TestChannelsDevicePollBehavior(t *testing.T) {
+	h := newDeviceHarness(t, mustChannelsSrc(t))
+	// begin 直查：user_code 透传、授权 URL、verifier 不回显
+	j := h.begin()
+	sess := j["session"].(map[string]interface{})
+	if sess["user_code"] != "ABCD-EFGH" {
+		t.Fatalf("begin 应透传 user_code, 实得 %v", sess["user_code"])
+	}
+	if j["url"] != "https://example/activate" {
+		t.Fatalf("begin 应给授权 URL, 实得 %v", j["url"])
+	}
+	if _, has := j["device_verifier"]; has {
+		t.Fatal("verifier 绝不能回显")
+	}
+	// 决策表全过
+	if !h.runDecisionTable() {
+		t.Fatal("device-poll 决策表未全过（见 runDecisionTable 内的行号）")
+	}
+}
+
+// mustChannelsSrc 读 bundled channels 源码。
+func mustChannelsSrc(t *testing.T) string {
+	t.Helper()
+	src, err := readBundledChannels(t)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return src
+}

internal/core/channels_more_test.go

@@ -0,0 +1,325 @@
+package core
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"llmsproxy/internal/lua"
+)
+
+// channelsMoreTest：新引入的三条真实流程 + 两个边界流的行为对齐测试。
+// 平台定义、决策表、端点逐字照抄参考（见 channels.lua 注释），测试照参考行为搬。
+
+// ---- Cline（WorkOS 设备码 + register 两步） ----
+
+// clineHarness 假上游：device code / authenticate / register 三端点。
+type clineHarness struct {
+	t    *testing.T
+	ps   *lua.Plugins
+	ts   *httptest.Server
+	tok  atomic.Value // 当前 authenticate 响应脚本 []scriptStep
+	step int32
+}
+
+func newClineHarness(t *testing.T) *clineHarness {
+	t.Helper()
+	h := &clineHarness{t: t}
+	h.tok.Store([]scriptStep{})
+	h.ts = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		case "/user_management/authorize/device":
+			fmt.Fprint(w, `{"device_code":"CD1","user_code":"WXYZ-1234","verification_uri":"https://example/activate","interval":1,"expires_in":600}`)
+		case "/user_management/authenticate":
+			i := int(atomic.AddInt32(&h.step, 1)) - 1
+			s := h.tok.Load().([]scriptStep)
+			if i >= len(s) {
+				w.WriteHeader(500)
+				fmt.Fprint(w, `{"error":"unexpected"}`)
+				return
+			}
+			if s[i].http != 0 && s[i].http != 200 {
+				w.WriteHeader(s[i].http)
+			}
+			fmt.Fprint(w, s[i].body)
+		case "/api/v1/auth/register":
+			// 校验请求体是驼峰 accessToken/refreshToken（参考 registerClineTokens）
+			b, _ := io.ReadAll(r.Body)
+			for _, want := range []string{`"accessToken":"WO-AT-1"`, `"refreshToken":"WO-RT-1"`} {
+				if !strings.Contains(string(b), want) {
+					t.Errorf("register 请求体缺 %s: %s", want, b)
+				}
+			}
+			fmt.Fprint(w, `{"success":true,"data":{"accessToken":"CLINE-AT-1","refreshToken":"CLINE-RT-1","userInfo":{"clineUserId":"u1","email":"dev@example.com"}}}`)
+		default:
+			w.WriteHeader(404)
+		}
+	}))
+	t.Cleanup(h.ts.Close)
+
+	c := kernelTestCore(t)
+	h.ps = c.Plugins()
+	src, _ := readBundledChannels(t)
+	if err := h.ps.LoadSource("channels", src); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := h.ps.Call("channels", "set_endpoints", map[string]interface{}{
+		"platform": "cline",
+		"device":   map[string]interface{}{"code_url": h.ts.URL + "/user_management/authorize/device", "token_url": h.ts.URL + "/user_management/authenticate"},
+		"register": map[string]interface{}{"url": h.ts.URL + "/api/v1/auth/register"},
+	}); err != nil {
+		t.Fatalf("set_endpoints: %v", err)
+	}
+	return h
+}
+
+func (h *clineHarness) begin() map[string]interface{} {
+	out, err := h.ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"cline"}`,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	return out.JSON.(map[string]interface{})
+}
+func (h *clineHarness) poll(id string) map[string]interface{} {
+	out, _ := h.ps.InvokeHandler("channels", "poll_login", &lua.HandlerRequest{
+		Method: "GET", Query: map[string]string{"id": id},
+	})
+	return out.JSON.(map[string]interface{})
+}
+
+// runWorkosTable 是 WorkOS（Cline）分支的行为契约，供变异测试杀 workos 变异体。
+// 决策表 5 行，跟 TestClineWorkOSBehavior 同源（这里只返回布尔，不报细节）。
+func (h *clineHarness) runWorkosTable() bool {
+	st := func(steps ...string) map[string]interface{} {
+		h.tok.Store(scriptSteps(steps...))
+		atomic.StoreInt32(&h.step, 0)
+		j := h.begin()
+		sid := j["session"].(map[string]interface{})["id"].(string)
+		return h.poll(sid)
+	}
+	_ = h.ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+	// 行1: 非200+error=authorization_pending → 等
+	if j := st(`400 {"error":"authorization_pending"}`); j["state"] != "pending" {
+		return false
+	}
+	// 行2: 非200+error=slow_down → 等 + interval+1（基线 1）
+	h.tok.Store(scriptSteps(`400 {"error":"slow_down"}`))
+	atomic.StoreInt32(&h.step, 0)
+	j := h.begin()
+	sid := j["session"].(map[string]interface{})["id"].(string)
+	if j2 := h.poll(sid); j2["state"] != "pending" {
+		return false
+	} else if iv, _ := j2["interval"].(float64); iv != 2 {
+		return false
+	}
+	// 行3: error=access_denied → failed
+	if j := st(`400 {"error":"access_denied","error_description":"user declined"}`); j["state"] != "failed" {
+		return false
+	}
+	// 行4: 成功（access+refresh）→ register 换 CLINE token → completed
+	if j := st(`{"access_token":"WO-AT-1","refresh_token":"WO-RT-1"}`); j["state"] != "completed" || j["account"] != "dev@example.com" {
+		return false
+	}
+	v, _, _ := h.ps.GetSecret("channels", "channels:cline")
+	if v != "CLINE-AT-1" {
+		return false
+	}
+	// 行5: 非200无error字段 → failed（不是死等）
+	if j := st(`500 {"foo":"bar"}`); j["state"] != "failed" {
+		return false
+	}
+	return true
+}
+
+// TestClineWorkOSBehavior：WorkOS 判据（error 字段）+ register 两步 + slow_down +1s。
+func TestClineWorkOSBehavior(t *testing.T) {
+	h := newClineHarness(t)
+	_ = h.ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+
+	st := func(steps ...string) map[string]interface{} {
+		h.tok.Store(scriptSteps(steps...))
+		atomic.StoreInt32(&h.step, 0)
+		j := h.begin()
+		sid := j["session"].(map[string]interface{})["id"].(string)
+		return h.poll(sid)
+	}
+
+	t.Run("行1 error=authorization_pending_非200也继续等", func(t *testing.T) {
+		if j := st(`400 {"error":"authorization_pending"}`); j["state"] != "pending" {
+			t.Fatalf("应 pending: %v", j)
+		}
+	})
+	t.Run("行2 error=slow_down_退避1秒", func(t *testing.T) {
+		h.tok.Store(scriptSteps(`400 {"error":"slow_down"}`))
+		atomic.StoreInt32(&h.step, 0)
+		j := h.begin()
+		sid := j["session"].(map[string]interface{})["id"].(string)
+		j2 := h.poll(sid)
+		if j2["state"] != "pending" {
+			t.Fatalf("应 pending: %v", j2)
+		}
+		if iv, _ := j2["interval"].(float64); iv != 2 { // 基线 1 + slow_down_step 1
+			t.Fatalf("WorkOS slow_down 应 interval+1s 得 2, 实得 %v", iv)
+		}
+	})
+	t.Run("行3 error=access_denied_失败带拒绝", func(t *testing.T) {
+		if j := st(`400 {"error":"access_denied","error_description":"user declined"}`); j["state"] != "failed" {
+			t.Fatalf("应 failed: %v", j)
+		}
+	})
+	t.Run("行4 成功_access+refresh_再打register_凭据是CLINE-AT-1", func(t *testing.T) {
+		j := st(`{"access_token":"WO-AT-1","refresh_token":"WO-RT-1"}`)
+		if j["state"] != "completed" || j["account"] != "dev@example.com" {
+			t.Fatalf("应 completed+dev@example.com: %v", j)
+		}
+		v, _, _ := h.ps.GetSecret("channels", "channels:cline")
+		if v != "CLINE-AT-1" {
+			t.Fatalf("凭据应是 register 的 CLINE-AT-1, 实得 %q", v)
+		}
+	})
+	t.Run("行5 非200无error字段_失败", func(t *testing.T) {
+		if j := st(`500 {"foo":"bar"}`); j["state"] != "failed" {
+			t.Fatalf("应 failed: %v", j)
+		}
+	})
+}
+
+// ---- Buddy（CodeBuddy state 轮询） ----
+
+type buddyHarness struct {
+	t       *testing.T
+	ps      *lua.Plugins
+	ts      *httptest.Server
+	tokStep int32
+	tokOk   int32
+	acctOk  int32
+}
+
+func newBuddyHarness(t *testing.T) *buddyHarness {
+	t.Helper()
+	h := &buddyHarness{t: t, tokOk: 1, acctOk: 1}
+	h.ts = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch {
+		case r.Method == "POST" && r.URL.Path == "/v2/plugin/auth/state":
+			fmt.Fprint(w, `{"data":{"state":"ST-1","authUrl":"https://www.codebuddy.cn/login/?platform=ide&state=ST-1"}}`)
+		case r.Method == "GET" && r.URL.Path == "/v2/plugin/auth/token":
+			i := atomic.AddInt32(&h.tokStep, 1)
+			if atomic.LoadInt32(&h.tokOk) == 0 {
+				w.WriteHeader(200)
+				fmt.Fprint(w, `{"code":11217,"message":"token not ready"}`)
+				return
+			}
+			if i == 1 {
+				fmt.Fprint(w, `{"code":11217,"message":"token not ready"}`)
+			} else {
+				fmt.Fprint(w, `{"data":{"accessToken":"BD-AT-1","refreshToken":"BD-RT-1"}}`)
+			}
+		case r.Method == "GET" && r.URL.Path == "/v2/plugin/login/account":
+			if atomic.LoadInt32(&h.acctOk) == 0 {
+				w.WriteHeader(500)
+				fmt.Fprint(w, `{"code":1}`)
+				return
+			}
+			fmt.Fprint(w, `{"data":{"nickName":"buddy-dev"}}`)
+		default:
+			w.WriteHeader(404)
+		}
+	}))
+	t.Cleanup(h.ts.Close)
+
+	c := kernelTestCore(t)
+	h.ps = c.Plugins()
+	src, _ := readBundledChannels(t)
+	if err := h.ps.LoadSource("channels", src); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := h.ps.Call("channels", "set_endpoints", map[string]interface{}{
+		"platform": "buddy",
+		"state":    map[string]interface{}{"state_url": h.ts.URL + "/v2/plugin/auth/state", "token_url": h.ts.URL + "/v2/plugin/auth/token", "account_url": h.ts.URL + "/v2/plugin/login/account"},
+	}); err != nil {
+		t.Fatalf("set_endpoints: %v", err)
+	}
+	return h
+}
+
+func (h *buddyHarness) begin() (map[string]interface{}, string) {
+	out, err := h.ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"buddy"}`,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	j := out.JSON.(map[string]interface{})
+	return j, j["session"].(map[string]interface{})["id"].(string)
+}
+func (h *buddyHarness) poll(id string) map[string]interface{} {
+	out, _ := h.ps.InvokeHandler("channels", "poll_login", &lua.HandlerRequest{
+		Method: "GET", Query: map[string]string{"id": id},
+	})
+	return out.JSON.(map[string]interface{})
+}
+
+// TestBuddyStatePollBehavior：state-poll 两拍（先 11217 后 token）+ 账号查询 + account 失败不阻塞。
+func TestBuddyStatePollBehavior(t *testing.T) {
+	h := newBuddyHarness(t)
+	_ = h.ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+
+	t.Run("行1 state申请给出authUrl", func(t *testing.T) {
+		j, _ := h.begin()
+		if j["url"] == "" {
+			t.Fatalf("应有 authUrl: %v", j)
+		}
+	})
+	t.Run("行2 首拍11217仍pending_次拍token_completed", func(t *testing.T) {
+		atomic.StoreInt32(&h.tokStep, 0)
+		_, sid := h.begin()
+		if j := h.poll(sid); j["state"] != "pending" {
+			t.Fatalf("首拍应 pending: %v", j)
+		}
+		if j := h.poll(sid); j["state"] != "completed" || j["account"] != "buddy-dev" {
+			t.Fatalf("次拍应 completed+buddy-dev: %v", j)
+		}
+		v, _, _ := h.ps.GetSecret("channels", "channels:buddy")
+		if v != "BD-AT-1" {
+			t.Fatalf("凭据应 BD-AT-1, 实得 %q", v)
+		}
+	})
+	t.Run("行3 account查询失败不阻塞_回退默认名", func(t *testing.T) {
+		atomic.StoreInt32(&h.tokStep, 0)
+		atomic.StoreInt32(&h.acctOk, 0)
+		defer atomic.StoreInt32(&h.acctOk, 1)
+		_, sid := h.begin()
+		h.poll(sid)
+		j := h.poll(sid)
+		if j["state"] != "completed" {
+			t.Fatalf("account 失败不应阻塞: %v", j)
+		}
+	})
+}
+
+// ---- 通用 ----
+
+func scriptSteps(steps ...string) []scriptStep {
+	var out []scriptStep
+	for _, b := range steps {
+		code := 200
+		switch {
+		case strings.HasPrefix(b, "400 "):
+			code, b = 400, b[4:]
+		case strings.HasPrefix(b, "403 "):
+			code, b = 403, b[4:]
+		case strings.HasPrefix(b, "500 "):
+			code, b = 500, b[4:]
+		}
+		out = append(out, scriptStep{http: code, body: b})
+	}
+	return out
+}

internal/core/channels_plugin_test.go

@@ -0,0 +1,241 @@
+package core
+
+import (
+	"encoding/json"
+	"testing"
+
+	"llmsproxy/internal/lua"
+)
+
+// TestChannelsPluginEndToEnd 是「同类插件」的完整验收：
+// 内置的 channels 插件在**真实内核**上跑通「登录 → 存凭据 → 接入为上游源」。
+//
+// 这条链路横跨本次所有新增能力，所以它是最有信息量的一个测试：
+//
+//	__llmsproxy_kernel  add_source 把平台注册为上游源
+//	内核权限            默认拒绝，需运维授权
+//	额外出站主机授权    登录域不在源列表里，默认被出站白名单拒绝（plugin.hosts）
+//	登录会话            插件自持（plugin.state.sessions），内核无登录概念
+func TestChannelsPluginEndToEnd(t *testing.T) {
+	c := kernelTestCore(t)
+	ps := c.Plugins()
+
+	src, err := lua.ReadBundledPlugin("channels")
+	if err != nil {
+		t.Fatalf("读取内置 channels 插件: %v", err)
+	}
+	if err := ps.LoadSource("channels", src); err != nil {
+		t.Fatalf("加载 channels: %v", err)
+	}
+
+	// 1. 平台清单可用（未授权任何东西也应能列出）
+	out, err := ps.InvokeHandler("channels", "list_platforms", &lua.HandlerRequest{
+		Method: "GET", Role: "admin",
+	})
+	if err != nil {
+		t.Fatalf("list_platforms: %v", err)
+	}
+	body, _ := out.JSON.(map[string]interface{})
+	plats, _ := body["platforms"].([]interface{})
+	if len(plats) < 3 {
+		t.Fatalf("应至少 3 个平台, 实得 %d", len(plats))
+	}
+
+	// 2. 发起登录 → 拿到会话与 URL
+	out, err = ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"example"}`, Role: "admin",
+	})
+	if err != nil {
+		t.Fatalf("start_login: %v", err)
+	}
+	m := out.JSON.(map[string]interface{})
+	if m["error"] != nil {
+		t.Fatalf("发起登录失败: %v", m["error"])
+	}
+	if m["url"] == nil || m["url"] == "" {
+		t.Fatal("应返回用户要打开的登录 URL")
+	}
+	sess := m["session"].(map[string]interface{})
+	id, _ := sess["id"].(string)
+	if id == "" {
+		t.Fatal("应返回会话 id")
+	}
+
+	// 3. 轮询到 pending
+	out, _ = ps.InvokeHandler("channels", "poll_login", &lua.HandlerRequest{
+		Method: "GET", Query: map[string]string{"id": id}, Role: "admin",
+	})
+	if st := out.JSON.(map[string]interface{})["state"]; st != "pending" {
+		t.Fatalf("应为 pending, 实得 %v", st)
+	}
+
+	// 4. 未授权内核动作时，接入源必须失败且**如实报告**（不是假装成功）
+	out, err = ps.InvokeHandler("channels", "complete_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"example","credential":"tok-1","account":"alice"}`, Role: "admin",
+	})
+	if err != nil {
+		t.Fatalf("complete_login: %v", err)
+	}
+	done := out.JSON.(map[string]interface{})
+	if done["source_error"] == nil {
+		t.Fatal("未授权 add_source 时应报告 source_error")
+	}
+	t.Logf("未授权时的如实报告: %v", done["source_error"])
+
+	// 凭据必须已在保险库（登录本身成功了）
+	tok, ok, _ := ps.GetSecret("channels", "channels:example")
+	if !ok || tok != "tok-1" {
+		t.Fatalf("凭据应已存保险库: ok=%v tok=%q", ok, tok)
+	}
+	// 且不得出现在 state 里
+	if stateHasSecret(ps, "channels", "tok-1") {
+		t.Fatal("★ 凭据泄漏进了 plugin.state")
+	}
+
+	// 5. 运维授权后，重新接入应当真的把源注册进内核
+	if err := ps.SetPluginKernelPermission("channels",
+		[]string{"add_source", "list_sources", "status"}, false); err != nil {
+		t.Fatal(err)
+	}
+	out, err = ps.InvokeHandler("channels", "complete_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"example","credential":"tok-1","account":"alice"}`, Role: "admin",
+	})
+	if err != nil {
+		t.Fatalf("complete_login(授权后): %v", err)
+	}
+	done = out.JSON.(map[string]interface{})
+	if done["source_error"] != nil {
+		t.Fatalf("授权后仍失败: %v", done["source_error"])
+	}
+	if done["source"] != "channels-example" {
+		t.Fatalf("源名不对: %v", done["source"])
+	}
+
+	// 内核侧必须真的有这个源
+	found := false
+	for _, s := range c.Sources() {
+		if s.Name == "channels-example" {
+			found = true
+			if s.BaseURL != "https://api.example.com/v1" {
+				t.Errorf("base_url 未落库: %q", s.BaseURL)
+			}
+		}
+	}
+	if !found {
+		t.Fatal("插件接入的源没有出现在内核源列表里")
+	}
+	// 平台声明的模型应可路由
+	models := c.Registry().ModelList()
+	hasModel := false
+	for _, id := range models {
+		if id == "example-chat" {
+			hasModel = true
+		}
+	}
+	if !hasModel {
+		t.Errorf("平台模型未进入可路由列表: %v", models)
+	}
+	t.Logf("已接入源 channels-example，模型可见: %v", models)
+}
+
+// TestChannelsPluginDeclaresItsNeeds 断言插件如实声明了它需要什么。
+//
+// 声明是运维审批的依据：插件不声明，设置页就无法显示"申请中"，
+// 运维只能等某个功能静默失效后才发现。
+func TestChannelsPluginDeclaresItsNeeds(t *testing.T) {
+	c := kernelTestCore(t)
+	ps := c.Plugins()
+	src, err := lua.ReadBundledPlugin("channels")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadSource("channels", src); err != nil {
+		t.Fatal(err)
+	}
+	// 内核动作声明
+	views := ps.KernelPermissions()
+	var declared []string
+	for _, v := range views {
+		if v.Plugin == "channels" {
+			declared = v.Declared
+		}
+	}
+	if len(declared) == 0 {
+		t.Fatal("channels 应声明它需要的内核动作")
+	}
+	for _, want := range []string{"add_source", "status"} {
+		found := false
+		for _, d := range declared {
+			if d == want {
+				found = true
+			}
+		}
+		if !found {
+			t.Errorf("应声明 %q, 实得 %v", want, declared)
+		}
+	}
+	// 额外出站主机声明（非源 host，需单独批准）
+	hosts := ps.PluginExtraHosts("channels")
+	if len(hosts) == 0 {
+		t.Fatal("channels 应声明额外出站主机")
+	}
+	t.Logf("声明内核动作=%v 额外主机=%v", declared, hosts)
+}
+
+// TestChannelsPluginHandlersRequireStatus 是一个契约回归：插件的每个 handler
+// 都必须返回 {status,json/body}，否则内核会**静默丢弃**它（实测踩过：
+// `return {url=…}` 得到 200 空响应，没有任何错误信息）。
+func TestChannelsPluginHandlersRequireStatus(t *testing.T) {
+	c := kernelTestCore(t)
+	ps := c.Plugins()
+	src, err := lua.ReadBundledPlugin("channels")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadSource("channels", src); err != nil {
+		t.Fatal(err)
+	}
+	_ = ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+
+	// 每个 handler 至少要在正常输入下给出非空响应
+	cases := []struct {
+		fn   string
+		req  *lua.HandlerRequest
+		what string
+	}{
+		{"list_platforms", &lua.HandlerRequest{Method: "GET"}, "平台列表"},
+		{"status_page", &lua.HandlerRequest{Method: "GET"}, "状态页"},
+		{"poll_login", &lua.HandlerRequest{Method: "GET", Query: map[string]string{"id": "nope"}}, "轮询(未知会话)"},
+		{"start_login", &lua.HandlerRequest{Method: "POST", Body: `{"platform":"example"}`}, "发起登录"},
+		{"start_login", &lua.HandlerRequest{Method: "POST", Body: `{"platform":"nope"}`}, "发起登录(未知平台)"},
+		{"complete_login", &lua.HandlerRequest{Method: "POST", Body: `{}`}, "完成登录(缺字段)"},
+		{"claim_daily", &lua.HandlerRequest{Method: "POST", Body: `{"platform":"example"}`}, "领额度(未登录)"},
+	}
+	for _, tc := range cases {
+		out, err := ps.InvokeHandler("channels", tc.fn, tc.req)
+		if err != nil {
+			t.Errorf("%s (%s) 返回错误: %v", tc.fn, tc.what, err)
+			continue
+		}
+		if out.JSON == nil && out.Body == "" {
+			t.Errorf("%s (%s) 返回空响应 —— 内核会静默丢弃，调用方看不到任何原因",
+				tc.fn, tc.what)
+		}
+	}
+}
+
+// jsonMarshalForTest 序列化一个接口值，用于检查凭据是否出现在 state 里。
+func jsonMarshalForTest(v interface{}) ([]byte, error) { return json.Marshal(v) }
+
+// stateHasSecret 检查凭据是否泄漏进 plugin.state。
+func stateHasSecret(ps *lua.Plugins, plugin, secret string) bool {
+	st := ps.State(plugin)
+	if st == nil {
+		return false
+	}
+	b, err := jsonMarshalForTest(st)
+	if err != nil {
+		return false
+	}
+	return contains(string(b), secret)
+}

internal/core/channels_pool_test.go

@@ -0,0 +1,322 @@
+package core
+
+import (
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/lua"
+)
+
+// channels_pool_test：对齐参考 account.* / model.* / provider.* 契约的行为测试。
+// 账号池增删改、上游源随 enabled/模型开关同步、供应商整体开关、备份导出。
+
+func poolCore(t *testing.T) (*Core, *lua.Plugins) {
+	t.Helper()
+	c := kernelTestCore(t)
+	ps := c.Plugins()
+	src, err := readBundledChannels(t)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadSource("channels", src); err != nil {
+		t.Fatal(err)
+	}
+	_ = ps.SetPluginKernelPermission("channels", []string{"add_source", "list_sources", "remove_source", "status", "probe_all", "reset_health"}, false)
+	return c, ps
+}
+
+// listSourcesViaKernel 直查内核 list_sources（actionListSources 返回 []map）。
+func listSourcesViaKernel(t *testing.T, c *Core) []map[string]interface{} {
+	t.Helper()
+	out, err := c.Invoke(nil, "list_sources", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("list_sources: %v", err)
+	}
+	if ms, ok := out.([]map[string]interface{}); ok {
+		return ms
+	}
+	raw, _ := out.([]interface{})
+	var srcs []map[string]interface{}
+	for _, s := range raw {
+		if m, ok := s.(map[string]interface{}); ok {
+			srcs = append(srcs, m)
+		}
+	}
+	return srcs
+}
+
+func invoke(t *testing.T, ps *lua.Plugins, fn, body string) map[string]interface{} {
+	t.Helper()
+	out, err := ps.InvokeHandler("channels", fn, &lua.HandlerRequest{Method: "POST", Body: body})
+	if err != nil {
+		t.Fatalf("invoke %s: %v", fn, err)
+	}
+	return out.JSON.(map[string]interface{})
+}
+
+// asArr 规整桥接后的数组。Lua 空表过桥是 map[string]interface{}（无法与空数组
+// 区分，见 vm.go luaTableToGoValue）；内核 list_sources 的 models 过桥是 []string。
+func asArr(v interface{}) []interface{} {
+	switch a := v.(type) {
+	case []interface{}:
+		return a
+	case []string:
+		out := make([]interface{}, len(a))
+		for i, s := range a {
+			out[i] = s
+		}
+		return out
+	}
+	return nil
+}
+
+func srcHasAccount(srcs []map[string]interface{}, needle string) (map[string]interface{}, bool) {
+	for _, sm := range srcs {
+		if name, _ := sm["name"].(string); strings.Contains(name, needle) {
+			return sm, true
+		}
+	}
+	return nil, false
+}
+
+// TestChannelsAccountPoolBehavior：账号池 CRUD + 上游源同步 + 模型/供应商开关 + 备份。
+func TestChannelsAccountPoolBehavior(t *testing.T) {
+	c, ps := poolCore(t)
+
+	t.Run("账号增删改", func(t *testing.T) {
+		m := invoke(t, ps, "account_add", `{"platform":"minimax","nickname":"我的号"}`)
+		if m["ok"] != true {
+			t.Fatalf("account_add: %v", m)
+		}
+		aid, _ := m["account_id"].(string)
+		if aid == "" {
+			t.Fatalf("account_add 应返回 account_id: %v", m)
+		}
+		m = invoke(t, ps, "account_list", `{"platform":"minimax"}`)
+		accts := asArr(m["accounts"])
+		if len(accts) != 1 {
+			t.Fatalf("应有1个账号: %d", len(accts))
+		}
+		a := accts[0].(map[string]interface{})
+		if a["nickname"] != "我的号" || a["enabled"] != true {
+			t.Fatalf("账号字段: %v", a)
+		}
+		invoke(t, ps, "account_update", `{"platform":"minimax","account_id":"`+aid+`","patch":{"nickname":"改名"}}`)
+		m = invoke(t, ps, "account_list", `{"platform":"minimax"}`)
+		if asArr(m["accounts"])[0].(map[string]interface{})["nickname"] != "改名" {
+			t.Fatalf("改名失败")
+		}
+		invoke(t, ps, "account_delete", `{"platform":"minimax","account_id":"`+aid+`"}`)
+		m = invoke(t, ps, "account_list", `{"platform":"minimax"}`)
+		if n := len(asArr(m["accounts"])); n != 0 {
+			t.Fatalf("删除后应为0, 实得 %d", n)
+		}
+	})
+
+	t.Run("模型开关同步上游源", func(t *testing.T) {
+		m := invoke(t, ps, "account_add", `{"platform":"minimax","credential":"MM-KEY"}`)
+		aid, _ := m["account_id"].(string)
+		sm, ok := srcHasAccount(listSourcesViaKernel(t, c), aid)
+		if !ok {
+			t.Fatal("带凭据账号应已同步源")
+		}
+		if n := len(asArr(sm["models"])); n != 3 {
+			t.Fatalf("源应有3模型, 实得 %d", n)
+		}
+		invoke(t, ps, "model_set_disabled", `{"platform":"minimax","model_id":"MiniMax-M3","disabled":true}`)
+		sm, _ = srcHasAccount(listSourcesViaKernel(t, c), aid)
+		if n := len(asArr(sm["models"])); n != 2 {
+			t.Fatalf("关1模型后源应2模型, 实得 %d", n)
+		}
+		invoke(t, ps, "model_set_all_disabled", `{"platform":"minimax","disabled":true}`)
+		m = invoke(t, ps, "provider_status", `{}`)
+		for _, s := range asArr(m["statuses"]) {
+			x := s.(map[string]interface{})
+			if x["id"] == "minimax" && x["closed"] != true {
+				t.Fatalf("全关模型后应 closed=true: %v", x)
+			}
+		}
+		invoke(t, ps, "model_set_all_disabled", `{"platform":"minimax","disabled":false}`)
+	})
+
+	t.Run("enabled开关同步上游源", func(t *testing.T) {
+		m := invoke(t, ps, "account_list", `{"platform":"minimax"}`)
+		accts := asArr(m["accounts"])
+		if len(accts) == 0 {
+			t.Fatal("上个用例应留有账号")
+		}
+		aid := accts[0].(map[string]interface{})["id"].(string)
+		before := len(listSourcesViaKernel(t, c))
+		invoke(t, ps, "account_update", `{"platform":"minimax","account_id":"`+aid+`","patch":{"enabled":false}}`)
+		if n := len(listSourcesViaKernel(t, c)); n != before-1 {
+			t.Fatalf("停用账号应删源: 前 %d 后 %d", before, n)
+		}
+		invoke(t, ps, "account_update", `{"platform":"minimax","account_id":"`+aid+`","patch":{"enabled":true}}`)
+		if n := len(listSourcesViaKernel(t, c)); n != before {
+			t.Fatalf("启用账号应重建源: 前 %d 后 %d", before, n)
+		}
+	})
+
+	t.Run("供应商整体开关", func(t *testing.T) {
+		invoke(t, ps, "provider_set_enabled", `{"platform":"minimax","enabled":false}`)
+		if _, ok := srcHasAccount(listSourcesViaKernel(t, c), "minimax"); ok {
+			t.Fatal("关闭渠道后不应有 minimax 源")
+		}
+		m := invoke(t, ps, "provider_status", `{}`)
+		for _, s := range asArr(m["statuses"]) {
+			x := s.(map[string]interface{})
+			if x["id"] == "minimax" && x["closed"] != true {
+				t.Fatalf("关闭后应 closed: %v", x)
+			}
+		}
+		invoke(t, ps, "provider_set_enabled", `{"platform":"minimax","enabled":true}`)
+		if _, ok := srcHasAccount(listSourcesViaKernel(t, c), "minimax"); !ok {
+			t.Fatal("开启渠道后应有 minimax 源")
+		}
+	})
+
+	t.Run("备份导出含凭据", func(t *testing.T) {
+		invoke(t, ps, "account_add", `{"platform":"cline","credential":"CL-KEY","nickname":"cl1"}`)
+		exp := invoke(t, ps, "backup_export", `{}`)
+		payload, _ := exp["payload"].(map[string]interface{})
+		if payload == nil {
+			t.Fatal("backup_export 应返回 payload")
+		}
+		creds, _ := payload["credentials"].(map[string]interface{})
+		found := false
+		for _, v := range creds {
+			if v == "CL-KEY" {
+				found = true
+			}
+		}
+		if !found {
+			t.Fatalf("备份应含凭据明文: %v", creds)
+		}
+	})
+}
+
+// TestChannelsExtendedRPCBehavior：对齐参考剩余 RPC 面（credits.status/balances、
+// account.reset/retest、model.setDisabledMany/clearDead、provider 排序、
+// gateway/backup.status、usage.badgePreference、onboarding）。
+func TestChannelsExtendedRPCBehavior(t *testing.T) {
+	_, ps := poolCore(t)
+
+	t.Run("credits_balances_无凭据报错_有凭据解析", func(t *testing.T) {
+		m := invoke(t, ps, "account_add", `{"platform":"minimax","nickname":"b1"}`)
+		aid, _ := m["account_id"].(string)
+		j := invoke(t, ps, "credits_balances", `{"platform":"minimax"}`)
+		var row map[string]interface{}
+		for _, a := range asArr(j["accounts"]) {
+			x := a.(map[string]interface{})
+			if x["accountId"] == aid {
+				row = x
+			}
+		}
+		if row == nil {
+			t.Fatal("balances 应含该账号")
+		}
+		if row["balance"] != nil || row["error"] == nil {
+			t.Fatalf("无凭据应 error 而非编造余额: %v", row)
+		}
+	})
+
+	t.Run("credits_status_无端点平台如实null", func(t *testing.T) {
+		m := invoke(t, ps, "account_add", `{"platform":"opencode","nickname":"oc1"}`)
+		aid, _ := m["account_id"].(string)
+		j := invoke(t, ps, "credits_status", `{"platform":"opencode"}`)
+		for _, a := range asArr(j["accounts"]) {
+			x := a.(map[string]interface{})
+			if x["accountId"] == aid && x["status"] != nil {
+				t.Fatalf("opencode 无状态端点应 null: %v", x)
+			}
+		}
+	})
+
+	t.Run("account_reset_clears", func(t *testing.T) {
+		invoke(t, ps, "account_add", `{"platform":"minimax","nickname":"r1"}`)
+		j := invoke(t, ps, "account_reset", `{"platform":"minimax","account_id":"__all__"}`)
+		if j["ok"] != true {
+			t.Fatalf("resetAll: %v", j)
+		}
+	})
+
+	t.Run("model_set_disabled_many", func(t *testing.T) {
+		invoke(t, ps, "model_set_disabled_many", `{"platform":"minimax","model_ids":["MiniMax-M3","MiniMax-M2.7-highspeed"],"disabled":true}`)
+		j := invoke(t, ps, "model_list", `{"platform":"minimax"}`)
+		off := 0
+		for _, x := range asArr(j["models"]) {
+			mm := x.(map[string]interface{})
+			if mm["disabled"] == true {
+				off++
+			}
+		}
+		if off != 2 {
+			t.Fatalf("批量关2模型后应2个disabled, 实得 %d", off)
+		}
+		invoke(t, ps, "model_set_disabled_many", `{"platform":"minimax","model_ids":["MiniMax-M3","MiniMax-M2.7-highspeed"],"disabled":false}`)
+	})
+
+	t.Run("model_clear_dead_只清黑名单里已下线的", func(t *testing.T) {
+		invoke(t, ps, "model_set_disabled", `{"platform":"minimax","model_id":"MiniMax-M3","disabled":true}`)
+		j := invoke(t, ps, "model_clear_dead", `{"platform":"minimax"}`)
+		if n := len(asArr(j["removed"])); n != 0 {
+			t.Fatalf("在册模型不算死模型, 实得 %d", n)
+		}
+		// 手动放一个不在 models 表里的黑名单项
+		invoke(t, ps, "model_set_disabled_many", `{"platform":"minimax","model_ids":["ghost-model"],"disabled":true}`)
+		j = invoke(t, ps, "model_clear_dead", `{"platform":"minimax"}`)
+		if n := len(asArr(j["removed"])); n != 1 {
+			t.Fatalf("死模型应被清: %v", j)
+		}
+		invoke(t, ps, "model_set_disabled", `{"platform":"minimax","model_id":"MiniMax-M3","disabled":false}`)
+	})
+
+	t.Run("provider_order", func(t *testing.T) {
+		invoke(t, ps, "provider_set_order", `{"order":["minimax","trae","bogus"]}`)
+		j := invoke(t, ps, "provider_get_order", `{}`)
+		order := asArr(j["order"])
+		if len(order) != 2 {
+			t.Fatalf("bogus 应被过滤, 实得 %v", order)
+		}
+		if order[0] != "minimax" {
+			t.Fatalf("顺序应保留: %v", order)
+		}
+	})
+
+	t.Run("backup_status", func(t *testing.T) {
+		invoke(t, ps, "account_add", `{"platform":"minimax","nickname":"s1"}`)
+		j := invoke(t, ps, "backup_status", `{}`)
+		if n, _ := j["accounts"].(float64); n < 1 {
+			t.Fatalf("backup_status 应统计账号: %v", j)
+		}
+	})
+
+	t.Run("badge_preference_非法值拒绝", func(t *testing.T) {
+		j := invoke(t, ps, "usage_badge_preference", `{"preference":"bogus"}`)
+		if j["error"] == nil {
+			t.Fatalf("非法偏好应拒绝: %v", j)
+		}
+		j = invoke(t, ps, "usage_badge_preference", `{"preference":"credits"}`)
+		if j["preference"] != "credits" {
+			t.Fatalf("合法偏好应保存: %v", j)
+		}
+	})
+
+	t.Run("onboarding_不支持平台如实报", func(t *testing.T) {
+		j := invoke(t, ps, "onboarding_status", `{"platform":"minimax"}`)
+		if j["supported"] != false {
+			t.Fatalf("minimax 无新人任务应 supported=false: %v", j)
+		}
+	})
+
+	t.Run("permanent_lock_读写", func(t *testing.T) {
+		j := invoke(t, ps, "credits_permanent_lock", `{"platform":"buddy","locked":true}`)
+		if j["locked"] != true {
+			t.Fatalf("锁定应生效: %v", j)
+		}
+		j = invoke(t, ps, "credits_permanent_lock", `{"platform":"buddy"}`)
+		if j["locked"] != true {
+			t.Fatalf("只读应回显锁定: %v", j)
+		}
+	})
+}

internal/core/channels_stress_test.go

@@ -0,0 +1,251 @@
+package core
+
+import (
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/lua"
+)
+
+// mustBox 构造一个真 SecretBox（与运行期同一实现，密钥落 runtime dir）。
+func mustBox(t *testing.T, runtimeFile string) *config.SecretBox {
+	t.Helper()
+	box, err := config.NewSecretBox(runtimeFile)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return box
+}
+
+// channelsCore 起一个带 channels 插件、可指定源码与 box 的核心（压力测试用）。
+func channelsCore(t *testing.T, src string) (*Core, *lua.Plugins) {
+	t.Helper()
+	c := kernelTestCore(t)
+	ps := c.Plugins()
+	if err := ps.LoadSource("channels", src); err != nil {
+		t.Fatalf("加载 channels: %v", err)
+	}
+	ps.SetSecretBox(mustBox(t, filepath.Join(t.TempDir(), "runtime.json")))
+	_ = ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+	return c, ps
+}
+
+// TestChannelsConcurrentLogins 压力：并发走 device-poll 登录。
+//
+// 插件是单 VM 单锁 —— 本测试验证的是**正确性下的并发**（会话互不串台、
+// 凭据互不可见、最终都 completed），真吞吐压力在 e2e 层（真二进制+HTTP 并发）。
+func TestChannelsConcurrentLogins(t *testing.T) {
+	var hits int32
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/oauth2/device/code":
+			n := atomic.AddInt32(&hits, 1)
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprintf(w, `{"device_code":"DEV-%d","user_code":"UC-%d","verification_uri":"https://example/activate","interval":1,"expires_in":600}`, n, n)
+		case "/oauth2/token":
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprint(w, `{"access_token":"TOK","refresh_token":"RF","nickname":"u"}`)
+		default:
+			w.WriteHeader(404)
+		}
+	}))
+	defer ts.Close()
+
+	_, ps := channelsCore(t, mustChannelsSrc(t))
+	// 把 minimax 端点指到假上游
+	if _, err := ps.Call("channels", "set_endpoints", map[string]interface{}{
+		"platform": "minimax",
+		"device": map[string]interface{}{
+			"code_url":  ts.URL + "/oauth2/device/code",
+			"token_url": ts.URL + "/oauth2/token",
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	const rounds = 12
+	var wg sync.WaitGroup
+	var failures int32
+	for i := 0; i < rounds; i++ {
+		wg.Add(1)
+		go func(n int) {
+			defer wg.Done()
+			out, err := ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+				Method: "POST", Body: `{"platform":"minimax"}`,
+			})
+			if err != nil {
+				t.Errorf("start_login #%d: %v", n, err)
+				atomic.AddInt32(&failures, 1)
+				return
+			}
+			j := out.JSON.(map[string]interface{})
+			sess, _ := j["session"].(map[string]interface{})
+			if sess == nil {
+				t.Errorf("start_login #%d 无会话: %v", n, j)
+				atomic.AddInt32(&failures, 1)
+				return
+			}
+			sid := sess["id"].(string)
+			for k := 0; k < 10; k++ {
+				out2, _ := ps.InvokeHandler("channels", "poll_login", &lua.HandlerRequest{
+					Method: "GET", Query: map[string]string{"id": sid},
+				})
+				m2 := out2.JSON.(map[string]interface{})
+				if m2["state"] == "completed" {
+					return
+				}
+				if m2["state"] == "failed" {
+					t.Errorf("会话 #%d failed: %v", n, m2)
+					atomic.AddInt32(&failures, 1)
+					return
+				}
+			}
+			t.Errorf("会话 #%d 10 次轮询未终态", n)
+			atomic.AddInt32(&failures, 1)
+		}(i)
+	}
+	wg.Wait()
+	if failures != 0 {
+		t.Fatalf("并发登录失败 %d 轮", failures)
+	}
+}
+
+// TestChannelsClaimStress 压力：同一平台并发领签到（幂等 + 无登录态丢失）。
+func TestChannelsClaimStress(t *testing.T) {
+	var claims int32
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/v1/checkin" {
+			atomic.AddInt32(&claims, 1)
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprint(w, `{"ok":true,"credits":10}`)
+			return
+		}
+		w.WriteHeader(404)
+		fmt.Fprint(w, `{}`)
+	}))
+	defer ts.Close()
+
+	_, ps := channelsCore(t, mustChannelsSrc(t))
+	if _, err := ps.Call("channels", "set_endpoints", map[string]interface{}{
+		"platform": "cookieish",
+		"checkin":  map[string]interface{}{"url": ts.URL + "/v1/checkin"},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.SetSecret("channels", "channels:cookieish", "PASTE-TOKEN"); err != nil {
+		t.Fatal(err)
+	}
+
+	const n = 30
+	var wg sync.WaitGroup
+	for i := 0; i < n; i++ {
+		wg.Add(1)
+		go func(k int) {
+			defer wg.Done()
+			out, err := ps.InvokeHandler("channels", "claim_daily", &lua.HandlerRequest{
+				Method: "POST", Body: `{"platform":"cookieish"}`,
+			})
+			if err != nil {
+				t.Errorf("claim #%d: %v", k, err)
+				return
+			}
+			j := out.JSON.(map[string]interface{})
+			if j["ok"] != true {
+				t.Errorf("claim #%d 未成功: %v", k, j)
+			}
+		}(i)
+	}
+	wg.Wait()
+	if got := atomic.LoadInt32(&claims); got != n {
+		t.Errorf("上游应收到 %d 次领取, 实得 %d", n, got)
+	}
+}
+
+// TestLegacyVaultPlaintextLoad 旧通路：老版本（无加密）写的明文库在新版本
+// （带 box）下必须能读 —— 凭据升级不丢。落盘则会在下次写时整体重封为密文。
+func TestLegacyVaultPlaintextLoad(t *testing.T) {
+	dir := t.TempDir()
+	ps := lua.NewPlugins(nil, dir)
+	if err := ps.SetSecret("legacy", "channels:old", "PLAIN-OLD-TOKEN"); err != nil {
+		t.Fatal(err)
+	}
+	ps.SetSecretBox(mustBox(t, filepath.Join(dir, "runtime.json")))
+	v, ok, err := ps.GetSecret("legacy", "channels:old")
+	if err != nil || !ok || v != "PLAIN-OLD-TOKEN" {
+		t.Fatalf("旧明文库读回失败: v=%q ok=%v err=%v", v, ok, err)
+	}
+	// 任意一次写触发整体重封
+	if err := ps.SetSecret("legacy", "k2", "NEW"); err != nil {
+		t.Fatal(err)
+	}
+	raw, err := os.ReadFile(filepath.Join(dir, ".secrets", "legacy.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(string(raw), "PLAIN-OLD-TOKEN") {
+		t.Fatalf("★ 重封后旧明文仍在盘上:\n%s", raw)
+	}
+	if !strings.Contains(string(raw), "enc:v1:") {
+		t.Fatalf("重封后应含密文:\n%s", raw)
+	}
+}
+
+// TestLegacySourceSealedConfig 旧通路：config.yaml 里明文 api_key 的旧配置，
+// 新版本加载后运行期拿到解密值、落盘被 seal（SealIfNeeded 语义）。
+func TestLegacySourceSealedConfig(t *testing.T) {
+	dir := t.TempDir()
+	cfgPath := filepath.Join(dir, "config.yaml")
+	yaml := "listen: 127.0.0.1:0\n" +
+		"gateway_keys:\n  - sk-legacy-0001\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"sources:\n" +
+		"  - name: legacy-src\n" +
+		"    base_url: https://legacy.example.com/v1\n" +
+		"    adapter: openai\n" +
+		"    api_key: PLAIN-LEGACY-KEY\n" +
+		"    max_concurrent: 2\n" +
+		"    models:\n" +
+		"      - id: m1\n" +
+		"        kind: chat\n" +
+		"        priority: 10\n"
+	if err := os.WriteFile(cfgPath, []byte(yaml), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := config.Load(cfgPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	c, err := NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+
+	got := c.Sources()
+	if len(got) != 1 || got[0].APIKey != "PLAIN-LEGACY-KEY" {
+		t.Fatalf("运行期 api_key 应解密: %#v", got)
+	}
+	raw, err := os.ReadFile(cfgPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(string(raw), "PLAIN-LEGACY-KEY") {
+		t.Fatalf("★ 旧明文 api_key 升级后仍明文落盘:\n%s", raw)
+	}
+	if !strings.Contains(string(raw), "enc:v1:") {
+		t.Fatalf("落盘应含 enc:v1 密文:\n%s", raw)
+	}
+	mk, err := os.ReadFile(filepath.Join(dir, "master.key"))
+	if err != nil || len(mk) != 64 {
+		t.Fatalf("master.key 应 64 hex, len=%d err=%v", len(mk), err)
+	}
+}

internal/core/channels_trae_test.go

@@ -0,0 +1,173 @@
+package core
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"llmsproxy/internal/lua"
+)
+
+// traeHarness：TRAE callback 流的测试台。假上游模拟 ExchangeToken；
+// 回调用 InvokeHandler 直打 oauth_callback（Query 表 = 平台重定向回来的参数）。
+type traeHarness struct {
+	t       *testing.T
+	ps      *lua.Plugins
+	ts      *httptest.Server
+	exchHit int32
+	exchOK  int32 // 1=成功 0=失败（atomic）
+}
+
+func newTraeHarness(t *testing.T) *traeHarness {
+	t.Helper()
+	h := &traeHarness{t: t, exchOK: 1}
+	h.ts = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/cloudide/api/v3/trae/oauth/ExchangeToken" {
+			w.WriteHeader(404)
+			return
+		}
+		n := atomic.AddInt32(&h.exchHit, 1)
+		w.Header().Set("Content-Type", "application/json")
+		if atomic.LoadInt32(&h.exchOK) == 0 {
+			w.WriteHeader(500)
+			fmt.Fprint(w, `{"error":"exchange down"}`)
+			return
+		}
+		// 校验请求体形状（对齐参考 exchangeTraeCallback 的请求构造）
+		body, _ := io.ReadAll(r.Body)
+		for _, want := range []string{`"ClientID"`, `"RefreshToken":"RT-FROM-CALLBACK"`, `"ClientSecret":"-"`} {
+			if !strings.Contains(string(body), want) {
+				t.Errorf("ExchangeToken 请求体缺 %s: %s", want, body)
+			}
+		}
+		fmt.Fprintf(w, `{"Token":"ACCESS-%d","RefreshToken":"RT-NEW-%d"}`, n, n)
+	}))
+	t.Cleanup(h.ts.Close)
+
+	c := kernelTestCore(t)
+	h.ps = c.Plugins()
+	src, err := readBundledChannels(t)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := h.ps.LoadSource("channels", src); err != nil {
+		t.Fatal(err)
+	}
+	_ = h.ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+	// ExchangeToken 指到假上游（set_endpoints 既是运维能力也是测试缝）
+	if _, err := h.ps.Call("channels", "set_endpoints", map[string]interface{}{
+		"platform": "trae",
+		"callback": map[string]interface{}{
+			"exchange_url": h.ts.URL + "/cloudide/api/v3/trae/oauth/ExchangeToken",
+		},
+	}); err != nil {
+		t.Fatalf("set_endpoints: %v", err)
+	}
+	return h
+}
+
+// beginAndCallback 发起登录（拿 nonce）后直接打 oauth_callback。
+// cbQuery 是回调的 query（不含 state，内部补）；值即平台重定向回来的原始参数。
+func (h *traeHarness) beginAndCallback(cbQuery map[string]string) (int, string) {
+	out, err := h.ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"trae"}`,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	j, ok := out.JSON.(map[string]interface{})
+	if !ok {
+		h.t.Fatalf("start_login 返回非对象: %#v", out.JSON)
+	}
+	sess := j["session"].(map[string]interface{})
+	nonce := sess["nonce"].(string)
+
+	q := map[string]string{"state": nonce}
+	for k, v := range cbQuery {
+		q[k] = v
+	}
+	res, err := h.ps.InvokeHandler("channels", "oauth_callback", &lua.HandlerRequest{
+		Method: "GET", Query: q,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	return res.Status, fmt.Sprintf("%v|%v", res.Body, res.JSON)
+}
+
+// TestTraeCallbackBehavior 是 TRAE callback 流的行为对齐测试。
+//
+// 决策表逐行搬自参考实现 parseTraeCallbackDetailed + exchangeTraeCallback
+// （trae-oauth.ts）。断言的是行为（凭据来源 / 错误文案 / 上游是否被调），不是实现细节。
+func TestTraeCallbackBehavior(t *testing.T) {
+	h := newTraeHarness(t)
+
+	t.Run("行1 refreshToken直传_ExchangeToken换新", func(t *testing.T) {
+		atomic.StoreInt32(&h.exchOK, 1)
+		st, body := h.beginAndCallback(map[string]string{
+			"refreshToken": "RT-FROM-CALLBACK",
+			"userInfo":     `{"UserID":"u1","ScreenName":"张三","TenantID":"t1"}`,
+		})
+		if st != 200 {
+			t.Fatalf("应 200: %d %s", st, body)
+		}
+		// 凭据必须是 ExchangeToken 的新 Token，不是回调原值
+		v, _, _ := h.ps.GetSecret("channels", "channels:trae")
+		if !strings.HasPrefix(v, "ACCESS-") {
+			t.Fatalf("凭据应是 ExchangeToken 的新 access, 实得 %q", v)
+		}
+	})
+
+	t.Run("行2 无refreshToken_userJwtToken兜底_不打ExchangeToken", func(t *testing.T) {
+		before := atomic.LoadInt32(&h.exchHit)
+		st, body := h.beginAndCallback(map[string]string{
+			"userJwt": `{"Token":"JWT-TOKEN-1","RefreshToken":""}`,
+		})
+		if st != 200 {
+			t.Fatalf("应 200: %d %s", st, body)
+		}
+		v, _, _ := h.ps.GetSecret("channels", "channels:trae")
+		if v != "JWT-TOKEN-1" {
+			t.Fatalf("兜底凭据应是 userJwt.Token, 实得 %q", v)
+		}
+		if atomic.LoadInt32(&h.exchHit) != before {
+			t.Fatal("无 refreshToken 不应调 ExchangeToken（参考实现分支2）")
+		}
+	})
+
+	t.Run("行3 只有code_PKCE流_明确报暂不支持", func(t *testing.T) {
+		st, body := h.beginAndCallback(map[string]string{"code": "PKCE-CODE-1"})
+		if st != 400 {
+			t.Fatalf("PKCE 流应 400: %d %s", st, body)
+		}
+		if !strings.Contains(body, "PKCE") {
+			t.Fatalf("错误应明确指向 PKCE 流, 实得 %s", body)
+		}
+	})
+
+	t.Run("行4 空回调_报缺失", func(t *testing.T) {
+		st, body := h.beginAndCallback(map[string]string{})
+		if st != 400 {
+			t.Fatalf("空回调应 400: %d %s", st, body)
+		}
+		if !strings.Contains(body, "refreshToken") {
+			t.Fatalf("错误应指出缺什么, 实得 %s", body)
+		}
+	})
+
+	t.Run("行5 ExchangeToken失败_透传阶段", func(t *testing.T) {
+		atomic.StoreInt32(&h.exchOK, 0)
+		defer atomic.StoreInt32(&h.exchOK, 1)
+		st, body := h.beginAndCallback(map[string]string{"refreshToken": "RT-FROM-CALLBACK"})
+		if st != 400 {
+			t.Fatalf("exchange 失败应 400: %d %s", st, body)
+		}
+		if !strings.Contains(body, "ExchangeToken") {
+			t.Fatalf("错误应指明 ExchangeToken 阶段, 实得 %s", body)
+		}
+	})
+}

internal/core/core.go

@@ -8,6 +8,7 @@ import (
 	"encoding/hex"
 	"fmt"
 	"log"
+	"net/url"
 	"os"
 	"path/filepath"
 	"sort"
@@ -37,6 +38,12 @@ type Core struct {
 	registry       *provider.Registry
 	autoChain      atomic.Pointer[scheduler.Chain] // chat AUTO chain
 	autoImageChain atomic.Pointer[scheduler.Chain] // image-generation AUTO chain
+
+	// keyAutoChains caches per-key AUTO chains, keyed by the gateway key.
+	// Entries are rebuilt by rebuildKeyAutoChains whenever sources or key
+	// config change; AutoChainFor reads them without taking c.mu so request
+	// handling stays lock-free like the global chain lookup.
+	keyAutoChains atomic.Pointer[map[string]*scheduler.Chain]
 }
 
 // New builds the core from a config file plus runtime overlay.
@@ -62,6 +69,34 @@ 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.
+	c.plugins.SetOutboundHosts(sourceHosts(cfg))
+	// Let the plugin layer validate access specs against the keys that exist.
+	// A stored allowlist naming a key that does not exist refuses every caller,
+	// which looks exactly like a broken endpoint.
+	lua.SetKeyNameLookup(func() []string {
+		// ListKeys is the structured view (bare `gateway_keys:` entries are
+		// expanded into records with names); cfg.GatewayKeys is still []string.
+		keys := c.ListKeys()
+		out := make([]string, 0, len(keys))
+		for _, k := range keys {
+			if k.Name != "" {
+				out = append(out, k.Name)
+			}
+		}
+		return out
+	})
+	// Operator overrides win over the plugin's own declared audience.
+	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)
 	}
@@ -96,6 +131,11 @@ 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)
@@ -257,6 +297,68 @@ func (c *Core) Registry() *provider.Registry { return c.registry }
 // AutoChain returns the current AUTO scheduling chain.
 func (c *Core) AutoChain() *scheduler.Chain { return c.autoChain.Load() }
 
+// AutoChainFor returns the AUTO chain to use for a request authenticated with
+// key. A key that declares its own chain gets that chain; every other key
+// inherits the global chain, so keys created before per-key chains existed
+// keep their current behaviour with no configuration change.
+//
+// It returns (chain, ok). ok is false only when the key has its own chain but
+// it compiled to nothing usable (e.g. every slot names a model that no longer
+// exists) — the caller must surface that rather than silently downgrading the
+// user to the global chain, which would be a confusing, invisible switch.
+func (c *Core) AutoChainFor(key string) (*scheduler.Chain, bool) {
+	if m := c.keyAutoChains.Load(); m != nil {
+		if ch, ok := (*m)[key]; ok {
+			return ch, true
+		}
+	}
+	return c.autoChain.Load(), true
+}
+
+// rebuildKeyAutoChains recompiles every per-key AUTO chain from the current
+// config and provider registry, then swaps the cache in one shot so requests
+// never observe a half-built map. Keys without their own chain are absent
+// from the map and fall back to the global chain.
+func (c *Core) rebuildKeyAutoChains() {
+	m := make(map[string]*scheduler.Chain)
+	for _, k := range c.cfg.Keys {
+		if !k.HasOwnAuto() {
+			continue
+		}
+		m[k.Key] = c.chainForRules(k.Auto, false)
+	}
+	c.keyAutoChains.Store(&m)
+}
+
+// SaveKeyAuto persists one key's own AUTO chain and recompiles the per-key
+// chain cache. Passing an empty list clears the override so the key inherits
+// the global chain again.
+func (c *Core) SaveKeyAuto(key string, entries []config.ModelScope) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	idx := -1
+	for i, k := range c.cfg.Keys {
+		if k.Key == key {
+			idx = i
+			break
+		}
+	}
+	if idx < 0 {
+		return fmt.Errorf("key not found")
+	}
+	if err := ValidateScopeQuotas(entries); err != nil {
+		return err
+	}
+	c.cfg.Keys[idx].Auto = cleanScopes(entries)
+	if err := c.saveConfig(); err != nil {
+		return err
+	}
+	// Rebuild the whole cache so a cleared override stops shadowing the
+	// global chain for this key.
+	c.rebuildKeyAutoChains()
+	return nil
+}
+
 // AutoImageChain returns the persisted image-generation AUTO chain snapshot.
 func (c *Core) AutoImageChain() *scheduler.Chain { return c.autoImageChain.Load() }
 
@@ -546,26 +648,98 @@ func (c *Core) rebuildRegistry() error {
 	}
 	c.buildAutoChain()
 	c.buildAutoImageChain()
+	c.rebuildKeyAutoChains()
+	// A source edit can add or remove an upstream host, which changes what a
+	// plugin is allowed to reach. Keeping this beside the chain rebuilds means
+	// one call site owns "config changed, refresh derived state".
+	c.plugins.SetOutboundHosts(sourceHosts(c.cfg))
 	return nil
 }
 
+// sourceHosts extracts the host of every configured source's base_url — the
+// 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
+		}
+		u, err := url.Parse(s.BaseURL)
+		if err != nil || u.Hostname() == "" {
+			continue
+		}
+		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
 // back to the source's best chat model and cooldown/quota bookkeeping would
 // key on a name that never matches.
 func (c *Core) buildAutoChain() {
+	c.autoChain.Store(c.chainForRules(c.cfg.Auto, false))
+}
+
+// chainForRules compiles a chat AUTO chain from a rule list. kindFilter is
+// false for the chat chain (image models are skipped) and true for the image
+// chain (only image models participate). Per-key chains always compile as
+// chat chains: a user's key configures the models *they* talk to.
+func (c *Core) chainForRules(rules []config.ModelScope, imageOnly bool) *scheduler.Chain {
 	prov := func(model, source string) scheduler.Provider {
 		p := c.registry.ProviderForSlot(model, source)
 		if p == nil {
 			return nil
 		}
-		if m := p.ModelByID(model); m != nil && m.Kind == "image" {
+		m := p.ModelByID(model)
+		if imageOnly && (m == nil || m.Kind != "image") {
+			return nil
+		}
+		if !imageOnly && m != nil && m.Kind == "image" {
 			return nil
 		}
 		return p
 	}
-	rules := c.cfg.Auto
 	sr := make([]scheduler.Rule, 0, len(rules))
 	for _, e := range rules {
 		model, source := e.Model, e.Source
@@ -574,7 +748,11 @@ func (c *Core) buildAutoChain() {
 				model = exact
 			}
 			if m := p.ModelByID(model); m != nil && m.Kind == "image" {
-				continue // image-kind models never join the chat AUTO chain
+				if !imageOnly {
+					continue // image-kind models never join the chat AUTO chain
+				}
+			} else if imageOnly {
+				continue
 			}
 			if source == "" {
 				source = p.Name()
@@ -589,7 +767,7 @@ func (c *Core) buildAutoChain() {
 			Hours:  e.Hours,
 		})
 	}
-	c.autoChain.Store(scheduler.BuildChain(sr, prov))
+	return scheduler.BuildChain(sr, prov)
 }
 
 // buildAutoImageChain rebuilds the image-generation AUTO chain snapshot.
@@ -706,6 +884,14 @@ 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
 	}
@@ -717,6 +903,13 @@ 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)

internal/core/device_harness_test.go

@@ -0,0 +1,180 @@
+package core
+
+import (
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"llmsproxy/internal/lua"
+)
+
+// readBundledChannels 读内置 channels 插件源（供行为/变异测试共用）。
+func readBundledChannels(t *testing.T) (string, error) {
+	t.Helper()
+	return lua.ReadBundledPlugin("channels")
+}
+
+// deviceHarness 是 device-poll 决策表的可复用测试台：假上游（脚本化逐次响应）
+// + channels 插件加载指定源码 + minimax 端点指到假上游。
+// 行为测试（TestChannelsDevicePollBehavior）与变异测试共用它 —— 决策表只有
+// 一份，两处不会漂移。
+type deviceHarness struct {
+	t    *testing.T
+	ps   *lua.Plugins
+	ts   *httptest.Server
+	step int32
+	// script 持续被 HTTP handler 原子读取
+	script atomic.Value // []scriptStep
+}
+
+type scriptStep struct {
+	http int
+	body string
+}
+
+func newDeviceHarness(t *testing.T, src string) *deviceHarness {
+	t.Helper()
+	h := &deviceHarness{t: t}
+	h.script.Store([]scriptStep{})
+
+	h.ts = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/oauth2/device/code":
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprint(w, `{"device_code":"DEV1","user_code":"ABCD-EFGH","verification_uri":"https://example/activate","interval":1,"expires_in":600}`)
+		case "/oauth2/token":
+			i := int(atomic.AddInt32(&h.step, 1)) - 1
+			script := h.script.Load().([]scriptStep)
+			w.Header().Set("Content-Type", "application/json")
+			if i >= len(script) {
+				w.WriteHeader(500)
+				fmt.Fprint(w, `{"error":"unexpected"}`)
+				return
+			}
+			if script[i].http != 0 && script[i].http != 200 {
+				w.WriteHeader(script[i].http)
+			}
+			fmt.Fprint(w, script[i].body)
+		default:
+			w.WriteHeader(404)
+		}
+	}))
+	t.Cleanup(h.ts.Close)
+
+	c := kernelTestCore(t)
+	h.ps = c.Plugins()
+	if err := h.ps.LoadSource("channels", src); err != nil {
+		t.Fatalf("加载: %v", err)
+	}
+	if _, err := h.ps.Call("channels", "set_endpoints", map[string]interface{}{
+		"platform": "minimax",
+		"device": map[string]interface{}{
+			"code_url":  h.ts.URL + "/oauth2/device/code",
+			"token_url": h.ts.URL + "/oauth2/token",
+		},
+	}); err != nil {
+		t.Fatalf("set_endpoints: %v", err)
+	}
+	return h
+}
+
+// setScript 设置 token 端点的逐次响应脚本。`400 {...}` 形式表示非 200。
+func (h *deviceHarness) setScript(steps ...string) {
+	var out []scriptStep
+	for _, b := range steps {
+		code := 200
+		switch {
+		case strings.HasPrefix(b, "400 "):
+			code, b = 400, b[4:]
+		case strings.HasPrefix(b, "403 "):
+			code, b = 403, b[4:]
+		}
+		out = append(out, scriptStep{http: code, body: b})
+	}
+	h.script.Store(out)
+	atomic.StoreInt32(&h.step, 0)
+}
+
+func (h *deviceHarness) begin() map[string]interface{} {
+	out, err := h.ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"minimax"}`,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	return out.JSON.(map[string]interface{})
+}
+
+func (h *deviceHarness) poll(id string) map[string]interface{} {
+	out, _ := h.ps.InvokeHandler("channels", "poll_login", &lua.HandlerRequest{
+		Method: "GET", Query: map[string]string{"id": id},
+	})
+	return out.JSON.(map[string]interface{})
+}
+
+// runDecisionTable 跑完整 8 行决策表，全部符合预期返回 true。
+// 这是行为契约的单一事实来源：行为测试用它断言通过；变异测试用它杀变异体。
+func (h *deviceHarness) runDecisionTable() bool {
+	ok := func(cond bool) bool { return cond }
+	st := func(steps ...string) map[string]interface{} {
+		h.setScript(steps...)
+		j := h.begin()
+		sid := j["session"].(map[string]interface{})["id"].(string)
+		return h.poll(sid)
+	}
+
+	// 行1: 200+pending → 等
+	if j := st(`{"status":"pending"}`); j["state"] != "pending" {
+		return false
+	}
+	// 行2: 200+slow_down → 等 + interval+5
+	h.setScript(`{"status":"slow_down"}`)
+	j := h.begin()
+	sid := j["session"].(map[string]interface{})["id"].(string)
+	if j2 := h.poll(sid); j2["state"] != "pending" {
+		return false
+	} else if iv, _ := j2["interval"].(float64); iv < 6 {
+		return false
+	}
+	// 行3: 200+denied → failed，message 区分于通用错误（杀"denied 不识别"变异体）
+	if j := st(`{"status":"denied"}`); j["state"] != "failed" {
+		return false
+	} else if msg, _ := j["message"].(string); !strings.Contains(msg, "用户拒绝") {
+		return false
+	}
+	// 行4: 200+expired_token → failed，message 带过期字样
+	if j := st(`{"status":"expired_token"}`); j["state"] != "failed" {
+		return false
+	} else if msg, _ := j["message"].(string); !strings.Contains(msg, "过期") {
+		return false
+	}
+	// 行5: 400+authorization_pending → 等
+	if j := st(`400 {"error":"authorization_pending"}`); j["state"] != "pending" {
+		return false
+	}
+	// 行6: 403+slow_down → 等 + interval+5
+	h.setScript(`403 {"error":"slow_down"}`)
+	j = h.begin()
+	sid = j["session"].(map[string]interface{})["id"].(string)
+	if j2 := h.poll(sid); j2["state"] != "pending" {
+		return false
+	} else if iv, _ := j2["interval"].(float64); iv < 6 {
+		return false
+	}
+	// 行7: 200+access_token → completed（需先授权 add_source）
+	_ = h.ps.SetPluginKernelPermission("channels", []string{"add_source", "status", "list_sources"}, false)
+	h.setScript(`{"access_token":"MM-ACCESS-1","refresh_token":"MM-RF-1","nickname":"mm-user"}`)
+	j = h.begin()
+	sid = j["session"].(map[string]interface{})["id"].(string)
+	if j2 := h.poll(sid); j2["state"] != "completed" || j2["account"] != "mm-user" {
+		return false
+	}
+	// 行8: 其它 error → failed 带文本
+	if j := st(`{"error":"server_error"}`); j["state"] != "failed" {
+		return false
+	}
+	return ok(true)
+}

internal/core/device_mutate_test.go

@@ -0,0 +1,73 @@
+package core
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestDevicePollMutants 是 device-poll 执行器的变异测试。
+//
+// 用户要求：移植要照参考实现的行为形成测试项，且要变异验证测试真的能杀掉
+// 「行为走样」的代码。每个用例把 channels 源里的一个决策分支做语义变异，
+// 断言决策表**会失败**（变异体被杀）。变异体幸存 = 该分支没有覆盖。
+func TestDevicePollMutants(t *testing.T) {
+	src, err := readBundledChannels(t)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// 锚点会命中 deviceOneStep 里两处同形代码（workos 分支 + minimax 分支），
+	// 全部替换：两个决策表（minimax + workos）都会杀，任一幸存即失败。
+	muts := []struct {
+		name string
+		from string
+		to   string
+	}{
+		{
+			name: "slow_down 不再加 interval",
+			from: `sess.device_interval = (sess.device_interval or d.interval or 5) + step`,
+			to:   `sess.device_interval = (sess.device_interval or d.interval or 5)`,
+		},
+		{
+			name: "denied 被当成继续等待",
+			from: `if resp.status == 200 and (status == "denied" or status == "access_denied") then`,
+			to:   `if resp.status == 200 and (status == "denied_x" or status == "access_denied") then`,
+		},
+		{
+			name: "expired_token 不被识别",
+			from: `if resp.status == 200 and (status == "expired" or status == "expired_token") then`,
+			to:   `if resp.status == 200 and (status == "expired" or status == "expired_token_x") then`,
+		},
+		{
+			name: "authorization_pending 被当成失败",
+			from: `if ecode == "authorization_pending" then return false end`,
+			to:   `if ecode == "authorization_pending_x" then return false end`,
+		},
+		{
+			name: "成功分支被禁用",
+			from: `if resp.status == 200 and rec.access_token then`,
+			to:   `if resp.status == 200 and rec.access_token and false then`,
+		},
+	}
+
+	for _, mu := range muts {
+		t.Run(mu.name, func(t *testing.T) {
+			if !strings.Contains(src, mu.from) {
+				t.Fatalf("锚点丢失，先更新实现与锚点：%q", mu.from)
+			}
+			// 锚点在 deviceOneStep 的 workos/minimax 两个分支各有一份同形代码，
+			// 全部替换：任一分支的变异体被任一决策表杀掉才算有覆盖。
+			mutated := strings.ReplaceAll(src, mu.from, mu.to)
+			if mutated == src {
+				t.Fatal("变异未生效")
+			}
+			// 基准必须全过（测试自身有效），变异体必须被杀（有覆盖）
+			if !newDeviceHarness(t, src).runDecisionTable() {
+				t.Fatal("基准决策表未全过 —— 测试自身失效")
+			}
+			if newDeviceHarness(t, mutated).runDecisionTable() {
+				t.Fatalf("★ 变异体幸存（%s）—— 该分支无覆盖", mu.name)
+			}
+		})
+	}
+}

internal/core/kernel_actions.go

@@ -0,0 +1,396 @@
+package core
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/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,
+		})
+	}
+	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 +1,336 @@
+package core
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/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()
+	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"),
+		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)
+	}
+	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"
+
+	"llmsproxy/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

@@ -0,0 +1,189 @@
+package core
+
+import (
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"llmsproxy/internal/config"
+)
+
+// AutoChainFor sits on the request hot path and reads keyAutoChains with no
+// lock, while SaveKeyAuto / rebuildRegistry replace that map from the admin
+// path. A "call it twice" test never interleaves those, so it would pass even
+// if the map were mutated in place. These tests run both sides at once under
+// -race and assert what readers actually observe.
+//
+// The property: a reader sees a chain that is either the key's own or the
+// global one — never a torn or empty one. If rebuildKeyAutoChains built the
+// map in place instead of swapping a fresh pointer, readers would sometimes
+// see a partially-filled map and silently fall back to the global chain, i.e.
+// a user quietly losing their configured chain under load.
+
+// concurrentChainConfig builds N keys, half with their own chain pointing at a
+// distinct model so a fallback to the global chain is distinguishable.
+func concurrentChainConfig(t *testing.T, n int) *config.Config {
+	t.Helper()
+	cfg := perKeyTestConfig(t)
+	cfg.Keys = nil
+	cfg.GatewayKeys = []string{"sk-gw-admin"}
+	cfg.Auto = []config.ModelScope{{Model: "gpt-4o", Source: "s1", Tier: 1}}
+	for i := 0; i < n; i++ {
+		k := config.GWKey{
+			Key:  fmt.Sprintf("sk-gw-k%03d", i),
+			Role: "user",
+			Name: fmt.Sprintf("user%d", i),
+		}
+		if i%2 == 0 {
+			k.Auto = []config.ModelScope{{Model: "gpt-4o-mini", Source: "s1", Tier: 1}}
+		}
+		cfg.Keys = append(cfg.Keys, k)
+	}
+	return cfg
+}
+
+func TestAutoChainForConcurrentWithSaveKeyAuto(t *testing.T) {
+	cfg := concurrentChainConfig(t, 40)
+	c := newTestCore(t, cfg)
+
+	var stop atomic.Bool
+	var wg sync.WaitGroup
+	var bad atomic.Int64
+
+	for r := 0; r < 8; r++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for !stop.Load() {
+				for i := 0; i < 40; i++ {
+					key := fmt.Sprintf("sk-gw-k%03d", i)
+					ch, ok := c.AutoChainFor(key)
+					if !ok || ch == nil || len(ch.Tiers) == 0 {
+						bad.Add(1)
+						continue
+					}
+					// Legal values are this key's own chain model (either
+					// side of the writer's flip) or the global chain's. The
+					// assertion is about the legal set, not a fixed value:
+					// the writers legitimately change the chain underneath.
+					got := ch.Tiers[0].Slots[0].Model
+					if got != "gpt-4o" && got != "gpt-4o-mini" {
+						bad.Add(1)
+					}
+				}
+			}
+		}()
+	}
+
+	for w := 0; w < 2; w++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			model := "gpt-4o-mini"
+			for i := 0; i < 25 && !stop.Load(); i++ {
+				for j := 0; j < 40; j += 2 {
+					if err := c.SaveKeyAuto(fmt.Sprintf("sk-gw-k%03d", j), []config.ModelScope{
+						{Model: model, Source: "s1", Tier: 1},
+					}); err != nil {
+						bad.Add(1)
+						return
+					}
+				}
+				if model == "gpt-4o-mini" {
+					model = "gpt-4o"
+				} else {
+					model = "gpt-4o-mini"
+				}
+			}
+		}()
+	}
+
+	time.Sleep(400 * time.Millisecond)
+	stop.Store(true)
+	wg.Wait()
+
+	if n := bad.Load(); n > 0 {
+		t.Fatalf("%d reads saw a chain outside {own, global} — the cache is "+
+			"being swapped in place instead of atomically", n)
+	}
+}
+
+// The registry rebuild path (AddSource / RemoveSource) also replaces the
+// cache, so readers must keep a usable chain throughout.
+//
+// Churned sources use a "churn" prefix on purpose: deleting a source a chain
+// references legitimately empties that chain (BuildChain drops unresolvable
+// slots), and asserting otherwise would be testing a fiction rather than the
+// code.
+func TestAutoChainForConcurrentWithSourceChanges(t *testing.T) {
+	cfg := concurrentChainConfig(t, 20)
+	c := newTestCore(t, cfg)
+
+	var stop atomic.Bool
+	var wg sync.WaitGroup
+	var bad atomic.Int64
+
+	for r := 0; r < 6; r++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for !stop.Load() {
+				for i := 0; i < 20; i++ {
+					ch, ok := c.AutoChainFor(fmt.Sprintf("sk-gw-k%03d", i))
+					if !ok || ch == nil || len(ch.Tiers) == 0 {
+						bad.Add(1)
+						continue
+					}
+					if got := ch.Tiers[0].Slots[0].Model; got != "gpt-4o" && got != "gpt-4o-mini" {
+						bad.Add(1)
+					}
+				}
+			}
+		}()
+	}
+
+	wg.Add(1)
+	go func() {
+		defer wg.Done()
+		for n := 0; n < 30 && !stop.Load(); n++ {
+			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}}},
+			)
+			_ = c.RemoveSource(name)
+		}
+	}()
+
+	time.Sleep(500 * time.Millisecond)
+	stop.Store(true)
+	wg.Wait()
+
+	if n := bad.Load(); n > 0 {
+		t.Fatalf("%d reads lost the chain while unrelated sources were churned", n)
+	}
+}
+
+// Many distinct keys: the lookup must stay a map hit and must not degrade into
+// returning the global chain as the key count grows.
+func TestAutoChainForManyKeys(t *testing.T) {
+	cfg := concurrentChainConfig(t, 500)
+	c := newTestCore(t, cfg)
+
+	for i := 0; i < 500; i++ {
+		key := fmt.Sprintf("sk-gw-k%03d", i)
+		ch, _ := c.AutoChainFor(key)
+		if ch == nil || len(ch.Tiers) == 0 {
+			t.Fatalf("key %s lost its chain with 500 keys configured", key)
+		}
+		want := "gpt-4o"
+		if i%2 == 0 {
+			want = "gpt-4o-mini"
+		}
+		if got := ch.Tiers[0].Slots[0].Model; got != want {
+			t.Fatalf("key %s got %q, want %q", key, got, want)
+		}
+	}
+}

internal/core/key_auto_test.go

@@ -0,0 +1,221 @@
+package core
+
+import (
+	"path/filepath"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/scheduler"
+)
+
+// Per-key AUTO chains let an admin give one user their own scheduling chain
+// while every other key keeps the global one. These tests pin the three
+// properties that make that safe:
+//
+//  1. a key WITH its own chain gets it;
+//  2. a key WITHOUT one inherits the global chain (backward compatible);
+//  3. the override is actually persisted, so it survives a restart.
+//
+// A test that only checked (1) would pass even if AutoChainFor ignored the key
+// and always returned the global chain whenever the two happened to be equal,
+// so (2) uses deliberately *different* chains.
+
+func perKeyTestConfig(t *testing.T) *config.Config {
+	t.Helper()
+	dir := t.TempDir()
+	return &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: "AUTO",
+		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}}},
+		},
+		// Global chain points at gpt-4o.
+		Auto: []config.ModelScope{{Model: "gpt-4o", Source: "s1", Tier: 1}},
+		Keys: []config.GWKey{
+			{Key: "sk-gw-admin", Role: "admin"},
+			{Key: "sk-gw-inherits", Role: "user", Name: "inherits"},
+			// This key's own chain points at a DIFFERENT model, so returning
+			// the global chain by mistake is detectable.
+			{Key: "sk-gw-own", Role: "user", Name: "own",
+				Auto: []config.ModelScope{{Model: "gpt-4o-mini", Source: "s1", Tier: 1}}},
+		},
+	}
+}
+
+func firstSlotModel(t *testing.T, c *Core, key string) string {
+	t.Helper()
+	ch, _ := c.AutoChainFor(key)
+	if ch == nil || len(ch.Tiers) == 0 {
+		t.Fatalf("key %s: no chain tiers", key)
+	}
+	slots := ch.Tiers[0].Slots
+	if len(slots) == 0 {
+		t.Fatalf("key %s: tier 1 has no slots", key)
+	}
+	return slots[0].Model
+}
+
+// A key that declares its own chain must be scheduled by it, not the global.
+func TestAutoChainForUsesKeyOwnChain(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	if got := firstSlotModel(t, c, "sk-gw-own"); got != "gpt-4o-mini" {
+		t.Fatalf("key with own chain must use it, got %q (want gpt-4o-mini)", got)
+	}
+	// Sanity: the global chain really is different, so the assertion above
+	// cannot pass by accident.
+	if got := firstSlotModel(t, c, "sk-gw-admin"); got != "gpt-4o" {
+		t.Fatalf("admin must use global chain, got %q (want gpt-4o)", got)
+	}
+}
+
+// A key with no own chain inherits the global one. This is what keeps every
+// pre-existing deployment working with no config change.
+func TestAutoChainForInheritsGlobalWhenUnset(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	if got := firstSlotModel(t, c, "sk-gw-inherits"); got != "gpt-4o" {
+		t.Fatalf("key without own chain must inherit global, got %q (want gpt-4o)", got)
+	}
+}
+
+// A key that is not in the key table at all (e.g. a seed key) must also
+// inherit the global chain rather than erroring or returning nil.
+func TestAutoChainForUnknownKeyInheritsGlobal(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	ch, ok := c.AutoChainFor("sk-gw-never-configured")
+	if !ok || ch == nil || len(ch.Tiers) == 0 {
+		t.Fatalf("unknown key must fall back to global chain, got ok=%v chain=%v", ok, ch)
+	}
+}
+
+// The per-key chain must survive a save/reload round trip: an admin configuring
+// a chain in the WebUI and restarting the gateway must not silently lose it.
+func TestSaveKeyAutoPersistsAndApplies(t *testing.T) {
+	cfg := perKeyTestConfig(t)
+	c := newTestCore(t, cfg)
+
+	if err := c.SaveKeyAuto("sk-gw-inherits", []config.ModelScope{
+		{Model: "gpt-4o-mini", Source: "s1", Tier: 1},
+	}); err != nil {
+		t.Fatalf("SaveKeyAuto: %v", err)
+	}
+	if got := firstSlotModel(t, c, "sk-gw-inherits"); got != "gpt-4o-mini" {
+		t.Fatalf("saved chain must apply immediately, got %q", got)
+	}
+
+	// Reload from disk exactly like a restart does.
+	reloaded, err := config.Load(cfg.Path)
+	if err != nil {
+		t.Fatalf("reload config: %v", err)
+	}
+	c2 := newTestCore(t, reloaded)
+	if got := firstSlotModel(t, c2, "sk-gw-inherits"); got != "gpt-4o-mini" {
+		t.Fatalf("per-key chain must survive restart, got %q", got)
+	}
+	// The other key must be unaffected by its neighbour's override.
+	if got := firstSlotModel(t, c2, "sk-gw-admin"); got != "gpt-4o" {
+		t.Fatalf("override must not leak to other keys, got %q", got)
+	}
+}
+
+// Clearing the override must return the key to the global chain, both in
+// memory and after a restart. This is what the WebUI's "use global" toggle
+// does, so a stale shadowing entry would silently pin the user to their old
+// chain forever.
+func TestSaveKeyAutoClearRestoresGlobal(t *testing.T) {
+	cfg := perKeyTestConfig(t)
+	c := newTestCore(t, cfg)
+
+	if err := c.SaveKeyAuto("sk-gw-own", nil); err != nil {
+		t.Fatalf("clear: %v", err)
+	}
+	if got := firstSlotModel(t, c, "sk-gw-own"); got != "gpt-4o" {
+		t.Fatalf("cleared key must inherit global again, got %q", got)
+	}
+
+	reloaded, err := config.Load(cfg.Path)
+	if err != nil {
+		t.Fatalf("reload: %v", err)
+	}
+	c2 := newTestCore(t, reloaded)
+	if got := firstSlotModel(t, c2, "sk-gw-own"); got != "gpt-4o" {
+		t.Fatalf("clear must survive restart, got %q", got)
+	}
+}
+
+// Saving an unknown key must fail loudly rather than creating a shadow entry
+// that no request can ever match.
+func TestSaveKeyAutoUnknownKeyFails(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	if err := c.SaveKeyAuto("sk-gw-nope", []config.ModelScope{{Model: "gpt-4o", Source: "s1"}}); err == nil {
+		t.Fatal("SaveKeyAuto on unknown key must return an error")
+	}
+}
+
+// Per-key quota validation must run, exactly like the model scope path: an
+// override is another place an operator can write a bad budget.
+func TestSaveKeyAutoRejectsBadQuota(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	err := c.SaveKeyAuto("sk-gw-inherits", []config.ModelScope{
+		{Model: "gpt-4o", Source: "s1", TokenQuota: -5},
+	})
+	if err == nil {
+		t.Fatal("negative token quota in a per-key auto chain must be rejected")
+	}
+}
+
+// Source changes must recompile per-key chains. A key whose chain names a
+// model that does not exist yet compiles to an empty chain (BuildChain drops
+// slots it cannot resolve — sending the request into a black hole would be
+// worse), and must start working as soon as the source appears.
+func TestKeyAutoChainsRebuildOnSourceChange(t *testing.T) {
+	cfg := perKeyTestConfig(t)
+	cfg.Keys = append(cfg.Keys, config.GWKey{
+		Key: "sk-gw-late", Role: "user", Name: "late",
+		Auto: []config.ModelScope{{Model: "new-model", Source: "s2", Tier: 1}},
+	})
+	c := newTestCore(t, cfg)
+
+	// s2 does not exist yet, so the chain has no usable slot.
+	if ch, _ := c.AutoChainFor("sk-gw-late"); chainSlotsOf(ch) != 0 {
+		t.Fatalf("chain naming an unknown model must compile empty, got %d slots",
+			chainSlotsOf(ch))
+	}
+	// The other keys are unaffected by the broken one.
+	if got := firstSlotModel(t, c, "sk-gw-admin"); got != "gpt-4o" {
+		t.Fatalf("broken per-key chain must not affect others, got %q", got)
+	}
+
+	// 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}},
+	})
+	if err := c.Reload(); err != nil {
+		t.Fatalf("Reload: %v", err)
+	}
+	if got := firstSlotModel(t, c, "sk-gw-late"); got != "new-model" {
+		t.Fatalf("per-key chain must resolve after the source is added, got %q", got)
+	}
+}
+
+// chainSlotsOf counts usable slots, tolerating a nil chain.
+func chainSlotsOf(ch *scheduler.Chain) int {
+	if ch == nil {
+		return 0
+	}
+	n := 0
+	for _, tn := range ch.Tiers {
+		n += len(tn.Slots)
+	}
+	return n
+}

internal/core/source_hosts_test.go

@@ -0,0 +1,108 @@
+package core
+
+import (
+	"net/url"
+	"testing"
+
+	"llmsproxy/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/gateway/api.go

@@ -445,6 +445,15 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 		out := g.stats.PeriodSnapshot(p, key, time.Now())
+		// The dashboard paints its records table from this payload's `records`
+		// field (paintRecords(st.records)), so a period view that omits the key
+		// renders an empty table no matter how much traffic the window had. The
+		// aggregate above is period-scoped; the record list is the same
+		// newest-first page the scroll handler pulls from /api/stats/records,
+		// which is deliberately not period-filtered (the cursor, not a date
+		// range, is its state) — that asymmetry is what the paging code
+		// already assumes.
+		page := g.stats.AuditPage("", limit, key)
 		writeJSON(w, http.StatusOK, map[string]interface{}{
 			"period":    out.Period,
 			"from":      out.From,
@@ -455,7 +464,21 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 			"by_status": out.Status,
 			"buckets":   out.Bucket,
 			"truncated": out.Truncated,
-			"key_names": g.keyNamesFor(),
+			// "Active requests" is a live gauge (requests in flight right now),
+			// not a period aggregate. Omitting it from the period response made
+			// the KPI read `st.active || 0` → 0 on every day/week/month view while
+			// 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(),
+			"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.
+			"next_cursor": page.Next,
+			"has_more":    page.HasMore,
+			"rotated":     page.Rotated,
+			"key_names":   g.keyNamesFor(),
 		})
 		return
 	}
@@ -567,6 +590,23 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 	}
 	snap := g.stats.Snapshot(limit, key)
 	snap["key_names"] = g.keyNamesFor()
+	// Records come from the audit files via the same AuditPage the scroll
+	// handler uses, not from Snapshot's in-memory ring. Two reasons:
+	//
+	//  - Ordering. The ring is oldest-first, AuditPage is newest-first. The
+	//    first screen and the paged continuation have to agree, or scrolling
+	//    splices new rows in above old ones.
+	//  - Depth. The ring holds maxRecs entries (a few hundred); the audit
+	//    files hold everything, so "all" actually means all and the cursor
+	//    can keep paging.
+	//
+	// It also carries next_cursor, without which the table can never load a
+	// second page.
+	page := g.stats.AuditPage("", limit, key)
+	snap["records"] = page.Records
+	snap["next_cursor"] = page.Next
+	snap["has_more"] = page.HasMore
+	snap["rotated"] = page.Rotated
 	writeJSON(w, http.StatusOK, snap)
 }
 

internal/gateway/apiv1.go

@@ -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/chat.go

@@ -21,21 +21,79 @@ import (
 )
 
 // 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"`
-	Messages        []types.ChatMessage    `json:"messages"`
-	Temperature     *float64               `json:"temperature,omitempty"`
-	MaxTokens       int                    `json:"max_tokens,omitempty"`
-	Stream          bool                   `json:"stream,omitempty"`
-	Tools           []interface{}          `json:"tools,omitempty"`
-	ToolChoice      interface{}            `json:"tool_choice,omitempty"`
+	Model       string              `json:"model"`
+	Messages    []types.ChatMessage `json:"messages"`
+	Temperature *float64            `json:"temperature,omitempty"`
+	MaxTokens   int                 `json:"max_tokens,omitempty"`
+	Stream      bool                `json:"stream,omitempty"`
+	Tools       []interface{}       `json:"tools,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"`
@@ -256,6 +314,16 @@ 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 quota being exhausted, and the message has to say which: telling an
+	// 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
+	// consulted, so this is the message that surfaces for AUTO.
+	if isAuto(model) {
+		return &quotaRejection{msg: fmt.Sprintf(
+			"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)}
+	}
 	return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
 }
 
@@ -400,7 +468,8 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 	// TestHooksFireOnRealDirectChat, not by reading the code.
 	g.fireStart(r.Context(), &req, "chat", model, len(req.Messages), len(req.Tools))
 	if isAuto(model) {
-		chain := g.core.AutoChain()
+		// The key's own chain wins; keys without one inherit the global chain.
+		chain, _ := g.core.AutoChainFor(reqKey(r.Context()))
 		if chain == nil || len(chain.Tiers) == 0 {
 			writeError(w, http.StatusServiceUnavailable, "no_provider", "no auto slot configured")
 			return
@@ -413,15 +482,15 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 		done := g.stats.Begin()
 		defer done()
 		inner := &types.ChatRequest{
-			Messages:        req.Messages,
-			Temperature:     req.Temperature,
-			MaxTokens:       req.MaxTokens,
-			Stream:          req.Stream,
-			Tools:           req.Tools,
-			ToolChoice:      req.ToolChoice,
-			DisableThinking: req.DisableThinking,
-			ExtraBody:       req.ExtraBody,
-			ClientSession:   clientSession,
+			Messages:      req.Messages,
+			Temperature:   req.Temperature,
+			MaxTokens:     req.MaxTokens,
+			Stream:        req.Stream,
+			Tools:         req.rawTools(),
+			ToolChoice:    req.rawToolChoice(),
+			Thinking:      req.thinkingFromClient(),
+			ExtraBody:     req.ExtraBody,
+			ClientSession: clientSession,
 		}
 		rec := &Req{
 			Key:  keyID(reqKey(ctx)),
@@ -451,6 +520,11 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 	}
+	// Note: a key whose scope excludes AUTO is rejected earlier, by
+	// checkQuota(ctx, "AUTO") above, which produces the same message. That
+	// ordering matters — checkQuota runs before the chain is consulted, so it
+	// is the check an AUTO request actually hits. Duplicating it here would be
+	// dead code.
 	cands, effective := g.resolveCands(r.Context(), &req)
 	if len(cands) == 0 {
 		writeError(w, http.StatusNotFound, "model_not_found", fmt.Sprintf("model %q is not configured", model))
@@ -468,16 +542,16 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 	defer done()
 
 	inner := &types.ChatRequest{
-		Model:           effective,
-		Messages:        req.Messages,
-		Temperature:     req.Temperature,
-		MaxTokens:       req.MaxTokens,
-		Stream:          req.Stream,
-		Tools:           req.Tools,
-		ToolChoice:      req.ToolChoice,
-		DisableThinking: req.DisableThinking,
-		ExtraBody:       req.ExtraBody,
-		ClientSession:   clientSession,
+		Model:         effective,
+		Messages:      req.Messages,
+		Temperature:   req.Temperature,
+		MaxTokens:     req.MaxTokens,
+		Stream:        req.Stream,
+		Tools:         req.rawTools(),
+		ToolChoice:    req.rawToolChoice(),
+		Thinking:      req.thinkingFromClient(),
+		ExtraBody:     req.ExtraBody,
+		ClientSession: clientSession,
 	}
 	rec := &Req{
 		Key:    keyID(reqKey(ctx)),
@@ -570,7 +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
@@ -773,6 +847,18 @@ func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedRespon
 	// Cache accounting: whichever source format the adapter normalized into
 	// (prompt_tokens_details.cached_tokens or legacy prompt_cache_hit_tokens),
 	// read it back so the request record carries the hit/miss split.
+	// reasoning_tokens is a SUBSET of completion_tokens and is billed at the
+	// completion rate by every provider measured, so it must NOT be added to
+	// Compl. Recording it separately is what lets a price rule split thinking
+	// from output for the providers that do price them differently, and what
+	// keeps the billing plugin from guessing.
+	//
+	// Leaving it unrecorded is worse than useless: the adapters already parse it
+	// 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)
+	}
 	if d := resp.TokenUsage.PromptTokensDetails; d != nil && d.CachedTokens > 0 {
 		rec.CacheHit = int64(d.CachedTokens)
 	}
@@ -789,29 +875,21 @@ func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedRespon
 	}
 }
 
-// writeChatCompletion renders a unified response as an OpenAI
-// chat.completion object. modelName is the id clients see as the serving
-// model: the requested id for direct routes, the exact slot model for AUTO.
-func writeChatCompletion(w http.ResponseWriter, resp *types.UnifiedResponse, modelName string) {
-	msg := RespMessage{Role: "assistant", Content: resp.Content}
-	if resp.ReasoningContent != "" {
-		msg.ReasoningContent = resp.ReasoningContent
+// writeChatCompletion renders a unified response through the out-bound
+// Serializer. The default serializer reproduces the OpenAI wire shape this
+// function used to build inline (locked byte-for-byte by e2e golden); an
+// adapter that declares serialize_response overrides it. modelName is the id
+// clients see as the serving model: the requested id for direct routes, the
+// exact slot model for AUTO.
+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 {
-		msg.ToolCalls = toolCallsWire(resp.ToolCalls)
-	}
-	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)
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	_, _ = w.Write(body)
 }
 
 // singleChat runs a direct (model-pinned) non-streaming request across the
@@ -836,7 +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
@@ -1028,9 +1106,12 @@ func (g *Gateway) fireEnd(rec *Req) {
 		// carried this (rec.CacheReported); the plugin payload did not, so any
 		// plugin-level cache accounting had to guess.
 		"cache_reported": rec.CacheReported,
-		"image_count":    rec.ImageCount,
-		"error":          rec.Err,
-		"time":           rec.Time,
+		// 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,
+		"error":            rec.Err,
+		"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
@@ -1085,24 +1166,29 @@ func mergeUsage(prev, cur *types.TokenUsage) *types.TokenUsage {
 	return &out
 }
 
-// pumpStream writes the full SSE sequence for a started stream: role
-// preamble, one chunk per unified delta, the terminating finish_reason, the
-// OpenAI-standard final usage chunk (empty choices) and [DONE]. modelName
-// follows writeChatCompletion's rule (requested id for direct routes, exact
-// slot model for AUTO).
-func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan types.UnifiedChunk, modelName string, t0 time.Time) {
+// pumpStream writes the full SSE sequence for a started stream. Every byte on
+// the wire comes from the Serializer, so the OpenAI frame shapes live in
+// serializer.go rather than here, and an adapter that declares serialize_chunk
+// replaces them wholesale.
+//
+// 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()
-	created := time.Now().Unix()
-	send := func(obj interface{}) bool {
-		b, err := json.Marshal(obj)
-		if err != nil {
-			return false
+	id := newWireID()
+	created := newCreated()
+	st := &StreamState{}
+	send := func(b []byte) bool {
+		if len(b) == 0 {
+			return true
 		}
 		if _, err := fmt.Fprintf(w, "data: %s\n\n", b); err != nil {
 			return false
@@ -1113,115 +1199,66 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 		return true
 	}
 
-	if !send(ChatChunk{
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{Role: "assistant"}}},
-	}) {
-		return
+	if b, err := ser.OpenRolePreamble(modelName, id, created, st); err == nil {
+		if !send(b) {
+			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{
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		b, err := ser.Chunk(&ck, modelName, id, created, st)
+		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 == "" {
-		finalFinish = "stop"
-	}
-	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
+
+	// Write the upstream's exact usage 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 lastUsage.Prompt > 0 {
-			rec.Prompt = int64(lastUsage.Prompt)
+	if st.Usage != nil {
+		if st.Usage.Prompt > 0 {
+			rec.Prompt = int64(st.Usage.Prompt)
 		}
-		if lastUsage.Completion > 0 {
-			rec.Compl = int64(lastUsage.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)
+		if st.Usage.Completion > 0 {
+			rec.Compl = int64(st.Usage.Completion)
 		}
+		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{
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
-			Choices: []ChunkChoice{},
-			Usage:   tut,
-			Cost:    lastCost,
-		})
-	} else if lastCost != "" {
-		// Cost arrived without any usage (OpenCode sends it on its own frame).
-		send(ChatChunk{
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
-			Choices: []ChunkChoice{},
-			Cost:    lastCost,
-		})
+
+	if frames, err := ser.Close(modelName, id, created, st); err == nil {
+		for _, b := range frames {
+			if !send(b) {
+				return
+			}
+		}
+	}
+	if marker := ser.DoneMarker(); marker != "" {
+		fmt.Fprintf(w, "data: %s\n\n", marker)
 	}
-	fmt.Fprintf(w, "data: [DONE]\n\n")
 	if flusher != nil {
 		flusher.Flush()
 	}
@@ -1252,7 +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
@@ -1285,7 +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
@@ -1315,7 +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) {
@@ -1428,3 +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_reasoning_test.go

@@ -0,0 +1,160 @@
+package gateway
+
+import (
+	"testing"
+
+	"llmsproxy/internal/types"
+)
+
+// TestRecordChatUsageKeepsReasoningTokens pins the reasoning-token plumbing.
+//
+// The adapters already parse completion_tokens_details.reasoning_tokens and
+// 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.
+//
+// The property that matters most: reasoning is a SUBSET of completion. Every
+// provider measured bills it at the completion rate, so a billing plugin that
+// added it to completion_tokens would double-charge every thinking token. The
+// test asserts completion is left alone.
+func TestRecordChatUsageKeepsReasoningTokens(t *testing.T) {
+	rec := &Req{}
+	resp := &types.UnifiedResponse{
+		TokenUsage: types.TokenUsage{
+			Prompt:     100,
+			Completion: 500,
+			CompletionTokensDetails: &types.CompletionTokensDetails{
+				ReasoningTokens: 320,
+			},
+		},
+	}
+	recordChatUsage(rec, &types.ChatRequest{}, resp)
+
+	if rec.Reasoning != 320 {
+		t.Errorf("Reasoning = %d, want 320", rec.Reasoning)
+	}
+	// The subset must not be folded into completion — it is already counted there.
+	if rec.Compl != 500 {
+		t.Errorf("Compl = %d, want 500 (reasoning is a subset, never additive)", rec.Compl)
+	}
+	if rec.Prompt != 100 {
+		t.Errorf("Prompt = %d, want 100", rec.Prompt)
+	}
+}
+
+// TestRecordChatUsageReasoningAbsentStaysZero: "not reported" and "reported as
+// zero" must both leave the field at 0, and neither may invent a number. A
+// plugin reading Reasoning cannot tell them apart, which is why the field is
+// documented as "0 means not reported" rather than "0 means no thinking".
+func TestRecordChatUsageReasoningAbsentStaysZero(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		u    types.TokenUsage
+	}{
+		{"no details object", types.TokenUsage{Prompt: 10, Completion: 20}},
+		{"details without reasoning", types.TokenUsage{
+			Prompt:                  10,
+			Completion:              20,
+			CompletionTokensDetails: &types.CompletionTokensDetails{},
+		}},
+	} {
+		rec := &Req{}
+		recordChatUsage(rec, &types.ChatRequest{}, &types.UnifiedResponse{TokenUsage: tc.u})
+		if rec.Reasoning != 0 {
+			t.Errorf("%s: Reasoning = %d, want 0", tc.name, rec.Reasoning)
+		}
+		if rec.Compl != 20 {
+			t.Errorf("%s: Compl = %d, want 20", tc.name, rec.Compl)
+		}
+	}
+}
+
+// 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/discover.go

@@ -0,0 +1,749 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"strings"
+	"time"
+
+	"llmsproxy/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, Priority: 50, Kind: "chat"})
+	}
+	for _, id := range image {
+		out = append(out, config.Model{ID: id, Priority: 70, Kind: "image"})
+	}
+	return out
+}

internal/gateway/discover_api.go

@@ -0,0 +1,403 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"llmsproxy/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"`
+}
+
+// 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)
+	if req.BaseURL == "" {
+		writeError(w, http.StatusBadRequest, "invalid_request", "base_url required")
+		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
+	}
+
+	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"
+
+	"llmsproxy/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 +1,342 @@
+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",
+			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 +1,648 @@
+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])
+	}
+	if got[2].Priority <= got[0].Priority {
+		t.Errorf("image model priority %d should sort after chat %d", got[2].Priority, got[0].Priority)
+	}
+}
+
+// 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/key_auto_api_test.go

@@ -0,0 +1,257 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// Per-key AUTO chain admin API. The behaviours pinned here are the ones that
+// are easy to get wrong and invisible when they are:
+//
+//   - /api/keys/me/auto must be reachable by a NON-admin (it is their own
+//     chain); routing it through the generic {key}/auto handler would 403 it.
+//   - /api/keys/{key}/auto must 403 a non-admin.
+//   - an empty PUT and a DELETE mean the same thing ("inherit global"), so the
+//     WebUI cannot persist a chain that would then 503 with no slots.
+//   - a PUT that resolves to zero slots is reported back, so an admin who
+//     mistyped a model learns it now instead of from the user's next 503.
+
+// keyAutoGateway builds a gateway with one admin key and one user key.
+func keyAutoGateway(t *testing.T) *Gateway {
+	t.Helper()
+	td := t.TempDir()
+	cfgPath := filepath.Join(td, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte("listen: :0"), 0644); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Path:         cfgPath,
+		AdapterDir:   filepath.Join(td, "adapters"),
+		RuntimeFile:  filepath.Join(td, "runtime.json"),
+		DefaultModel: "AUTO",
+		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}}},
+		},
+	}
+	if err := cfg.ApplyDefaults(); err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	g, err := New(c)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g
+}
+
+// doReqAs issues an authenticated request as a specific key.
+func doReqAs(t *testing.T, g *Gateway, key, method, path, body string) *httptest.ResponseRecorder {
+	t.Helper()
+	req, _ := http.NewRequest(method, path, strings.NewReader(body))
+	req.Header.Set("Authorization", "Bearer "+key)
+	if body != "" {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	rr := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rr, req)
+	return rr
+}
+
+func TestKeyAutoAPIPermissions(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	// A user may read their own chain...
+	if rr := doReqAs(t, g, "sk-user", "GET", "/api/keys/me/auto", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys/me/auto as user = %d, want 200 (body %s)", rr.Code, rr.Body.String())
+	}
+	// ...but not another key's.
+	if rr := doReqAs(t, g, "sk-user", "GET", "/api/keys/sk-admin/auto", ""); rr.Code != http.StatusForbidden {
+		t.Fatalf("GET other key's auto as user = %d, want 403", rr.Code)
+	}
+	// ...and not their own via the admin route either.
+	if rr := doReqAs(t, g, "sk-user", "PUT", "/api/keys/sk-user/auto", `{"auto":[{"model":"gpt-4o"}]}`); rr.Code != http.StatusForbidden {
+		t.Fatalf("user PUT own auto = %d, want 403 (only admin configures)", rr.Code)
+	}
+}
+
+func TestKeyAutoAPICrudAndInherit(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	// Initially inherits.
+	var got struct {
+		Inherits bool                `json:"inherits"`
+		Auto     []config.ModelScope `json:"auto"`
+	}
+	rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if rr.Code != http.StatusOK {
+		t.Fatalf("GET = %d: %s", rr.Code, rr.Body.String())
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits || len(got.Auto) != 0 {
+		t.Fatalf("fresh key must inherit global, got inherits=%v auto=%v", got.Inherits, got.Auto)
+	}
+
+	// Admin sets a chain; the response reports the resolved slot count.
+	rr = doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"gpt-4o-mini","source":"s1","tier":1}]}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	var put struct {
+		Inherits bool `json:"inherits"`
+		Slots    int  `json:"slots"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &put); err != nil {
+		t.Fatal(err)
+	}
+	if put.Inherits {
+		t.Fatal("after a non-empty PUT the key must no longer inherit")
+	}
+	if put.Slots != 1 {
+		t.Fatalf("resolved slots = %d, want 1", put.Slots)
+	}
+
+	// And it reads back as an own chain.
+	rr = doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if got.Inherits || len(got.Auto) != 1 || got.Auto[0].Model != "gpt-4o-mini" {
+		t.Fatalf("own chain must read back, got inherits=%v auto=%v", got.Inherits, got.Auto)
+	}
+
+	// DELETE restores inheritance.
+	if rr := doReqAs(t, g, "sk-admin", "DELETE", "/api/keys/sk-user/auto", ""); rr.Code != http.StatusOK {
+		t.Fatalf("DELETE = %d: %s", rr.Code, rr.Body.String())
+	}
+	rr = doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits {
+		t.Fatal("after DELETE the key must inherit the global chain again")
+	}
+}
+
+// An empty PUT must behave like DELETE. Otherwise a WebUI that submits an
+// empty list (every slot deleted in the editor) would persist an empty chain
+// and the next AUTO request would fail with no_provider.
+func TestKeyAutoAPIEmptyPutClearsOverride(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	if rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"gpt-4o","source":"s1","tier":1}]}`); rr.Code != http.StatusOK {
+		t.Fatalf("seed PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto", `{"auto":[]}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("empty PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	var got struct {
+		Inherits bool `json:"inherits"`
+	}
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", ""); rr.Code != http.StatusOK {
+		t.Fatal(rr.Body.String())
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits {
+		t.Fatal("an empty PUT must clear the override, not persist an empty chain")
+	}
+}
+
+// A chain naming a model that does not exist compiles to zero slots. The API
+// must say so at write time instead of letting the user's next request 503.
+func TestKeyAutoAPIReportsUnresolvableSlots(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"does-not-exist","source":"s1","tier":1}]}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	var put struct {
+		Inherits bool `json:"inherits"`
+		Slots    int  `json:"slots"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &put); err != nil {
+		t.Fatal(err)
+	}
+	if put.Slots != 0 {
+		t.Fatalf("unknown model must resolve to 0 slots, got %d", put.Slots)
+	}
+}
+
+func TestKeyAutoAPIUnknownKey404(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-nope/auto", ""); rr.Code != http.StatusNotFound {
+		t.Fatalf("GET unknown key auto = %d, want 404", rr.Code)
+	}
+	if rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-nope/auto", `{"auto":[{"model":"gpt-4o"}]}`); rr.Code != http.StatusNotFound {
+		t.Fatalf("PUT unknown key auto = %d, want 404", rr.Code)
+	}
+}
+
+// A bad quota must be rejected with 400, not persisted and not silently
+// clamped — same rule as the model scope path.
+func TestKeyAutoAPIRejectsBadQuota(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"gpt-4o","source":"s1","tier":1,"token_quota":-1}]}`)
+	if rr.Code != http.StatusBadRequest {
+		t.Fatalf("negative quota PUT = %d, want 400 (body %s)", rr.Code, rr.Body.String())
+	}
+	// Nothing may have been persisted.
+	var got struct {
+		Inherits bool `json:"inherits"`
+	}
+	get := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if get.Code != http.StatusOK {
+		t.Fatal(get.Body.String())
+	}
+	if err := json.Unmarshal(get.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits {
+		t.Fatal("a rejected PUT must not persist a chain")
+	}
+}
+
+// The other key endpoints must keep working: the {key}/auto route must not
+// shadow /api/keys/{key} itself.
+func TestKeyAutoAPIDoesNotShadowKeyRoutes(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys = %d, want 200", rr.Code)
+	}
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/me", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys/me = %d, want 200", rr.Code)
+	}
+	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/keys.go

@@ -7,6 +7,7 @@ import (
 	"strings"
 
 	"llmsproxy/internal/config"
+	"llmsproxy/internal/scheduler"
 )
 
 // handleKeysAPI manages gateway keys: GET /api/keys (admin: all keys),
@@ -17,10 +18,22 @@ func (g *Gateway) handleKeysAPI(w http.ResponseWriter, r *http.Request) {
 	path = strings.Trim(path, "/")
 	role := reqRole(r.Context())
 
+	// /api/keys/me/auto — a user's own chain, readable without admin rights.
+	// Must be matched before the generic {key}/auto case below so it does not
+	// ask admin rights of a plain user asking about their own chain.
+	if path == "me/auto" {
+		g.handleKeyMeAuto(w, r)
+		return
+	}
 	if path == "me" {
 		g.handleKeyMe(w, r)
 		return
 	}
+	// /api/keys/{key}/auto — the per-key AUTO chain editor.
+	if i := strings.LastIndex(path, "/auto"); i > 0 && path[i+len("/auto"):] == "" {
+		g.handleKeyAutoAPI(w, r, path[:i])
+		return
+	}
 	if role != "admin" {
 		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
 		return
@@ -139,6 +152,101 @@ func (g *Gateway) allowedModels(ctx context.Context) []config.ModelScope {
 	return rec.Models
 }
 
+// handleKeyAutoAPI manages one key's own AUTO chain:
+// GET /api/keys/{key}/auto returns it, PUT replaces it, DELETE clears the
+// override so the key falls back to the global chain. Admin only — a user can
+// inspect their own chain through GET /api/keys/me/auto.
+//
+// An empty chain and a cleared chain are deliberately the same state: "inherit
+// the global chain". That keeps the WebUI's "use global" toggle a plain DELETE
+// with no sentinel value to carry through config.yaml.
+func (g *Gateway) handleKeyAutoAPI(w http.ResponseWriter, r *http.Request, key string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	rec, ok := g.core.FindKey(key)
+	if !ok {
+		writeError(w, http.StatusNotFound, "not_found", "key not found")
+		return
+	}
+	switch r.Method {
+	case http.MethodGet:
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"key":      key,
+			"auto":     rec.Auto,
+			"inherits": !rec.HasOwnAuto(),
+		})
+	case http.MethodPut, http.MethodPatch:
+		var body struct {
+			Auto []config.ModelScope `json:"auto"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+			return
+		}
+		if len(body.Auto) == 0 {
+			// Treat an empty PUT as "give up the override" so the endpoint
+			// cannot persist a chain that would 503 with no slots.
+			if err := g.core.SaveKeyAuto(key, nil); err != nil {
+				writeError(w, http.StatusBadRequest, "key_error", err.Error())
+				return
+			}
+			writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "inherits": true})
+			return
+		}
+		if err := g.core.SaveKeyAuto(key, body.Auto); err != nil {
+			writeError(w, http.StatusBadRequest, "key_error", err.Error())
+			return
+		}
+		// Report whether the chain actually resolved to usable slots. An
+		// admin who typed a model that no longer exists should learn it here,
+		// not from the user's next 503.
+		chain, _ := g.core.AutoChainFor(key)
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"ok": true, "inherits": false, "slots": chainSlots(chain),
+		})
+	case http.MethodDelete:
+		if err := g.core.SaveKeyAuto(key, nil); err != nil {
+			writeError(w, http.StatusBadRequest, "key_error", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "inherits": true})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+	}
+}
+
+// chainSlots counts the usable slots in a chain, for API feedback after an
+// edit. A chain that compiled to zero slots is reported so the caller can warn
+// instead of letting the next request fail with no_provider.
+func chainSlots(ch *scheduler.Chain) int {
+	if ch == nil {
+		return 0
+	}
+	n := 0
+	for _, tn := range ch.Tiers {
+		n += len(tn.Slots)
+	}
+	return n
+}
+
+// handleKeyMeAuto lets a user read their own AUTO chain without admin rights.
+func (g *Gateway) handleKeyMeAuto(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	rec, ok := g.core.FindKey(reqKey(r.Context()))
+	if !ok {
+		writeError(w, http.StatusUnauthorized, "invalid_api_key", "key not found")
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{
+		"key": rec.Key, "auto": rec.Auto, "inherits": !rec.HasOwnAuto(),
+	})
+}
+
 // handleAutoAPI manages the AUTO scheduling slots: GET /api/auto returns the
 // current rules; PUT /api/auto replaces them (admin only).
 func (g *Gateway) handleAutoAPI(w http.ResponseWriter, r *http.Request) {

internal/gateway/messages.go

@@ -0,0 +1,379 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"llmsproxy/internal/provider"
+	"llmsproxy/internal/scheduler"
+	"llmsproxy/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"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/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", Priority: 100}},
+		}},
+	}
+
+	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"
+
+	"llmsproxy/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"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/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/plugin_call_api_test.go

@@ -0,0 +1,150 @@
+package gateway
+
+// The /api/plugins/{name}/call/{fn} endpoint splits its path with string
+// surgery, and the split point is easy to get wrong in a way that still
+// "works": keeping the "call/" segment in the function name produced
+// `does not export "call/collect_account"` — the plugin was correct, the router
+// was not, and every call 502'd while the exports list looked perfectly fine.
+//
+// These tests drive the real router, so a refactor that reintroduces the
+// off-by-one is caught here rather than in production.
+
+import (
+	"encoding/json"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// callPlugin exposes exactly one export plus a function that is NOT exported.
+// If the exports gate is not honoured, secretFn is reachable and with it the
+// plugin's private state — for a plugin holding an API key that means handing
+// the key to anyone who can POST to the gateway.
+const callPlugin = `
+local P = {}
+P.name = "calltest"
+P.version = "1.0.0"
+P.exports = { "add" }
+P.secret = "do-not-leak"
+function P.add(a)
+  return { result = (tonumber(a and a.x) or 0) + 1 }
+end
+function P.secretFn()
+  return { leaked = P.secret }
+end
+return P
+`
+
+func gatewayWithCallPlugin(t *testing.T) *Gateway {
+	t.Helper()
+	dir := t.TempDir()
+	cfgPath := filepath.Join(dir, "config.yaml")
+	body := "listen: :0\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"plugin_dir: " + filepath.Join(dir, "plugins") + "\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"gateway_keys:\n  - sk-test\n"
+	if err := os.WriteFile(cfgPath, []byte(body), 0600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := config.Load(cfgPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	if err := c.Plugins().LoadSource("calltest", callPlugin); err != nil {
+		t.Fatalf("load calltest: %v", err)
+	}
+	g, err := New(c)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g
+}
+
+func TestPluginCallSplitsFunctionNameCorrectly(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodPost,
+		"/api/plugins/calltest/call/add", `{"x":41}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("call failed: HTTP %d %s", rr.Code, rr.Body.String())
+	}
+	var out struct {
+		Result struct {
+			Result int `json:"result"`
+		} `json:"result"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode: %v (body=%s)", err, rr.Body.String())
+	}
+	// 41 + 1 == 42. Reaching 42 proves BOTH halves of the contract: the "/call/"
+	// segment did not leak into the function name, and the payload arrived.
+	if out.Result.Result != 42 {
+		t.Fatalf("got %d, want 42 — either the /call/ segment leaked into the "+
+			"function name or the argument table never reached Lua", out.Result.Result)
+	}
+}
+
+func TestPluginCallRefusesUndeclaredFunction(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodPost,
+		"/api/plugins/calltest/call/secretFn", `{}`)
+	if rr.Code == http.StatusOK {
+		t.Fatalf("an unexported function must not be callable, got 200: %s",
+			rr.Body.String())
+	}
+	if strings.Contains(rr.Body.String(), "do-not-leak") {
+		t.Fatal("the refusal leaked the plugin's private state")
+	}
+}
+
+// A plugin that declares no exports at all has no callable surface. Reading the
+// missing list as "allow everything" would turn every plugin into a remote
+// evaluator the moment this endpoint ships.
+func TestPluginCallWithoutExportsHasNoCallableSurface(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	const bare = `
+local P = {}
+P.name = "bare"
+P.version = "1.0.0"
+function P.anything()
+  return { ok = true }
+end
+return P
+`
+	if err := g.core.Plugins().LoadSource("bare", bare); err != nil {
+		t.Fatalf("load bare: %v", err)
+	}
+	rr := doReq(t, g, http.MethodPost,
+		"/api/plugins/bare/call/anything", `{}`)
+	if rr.Code == http.StatusOK {
+		t.Fatalf("a plugin with no plugin.exports must expose nothing, got 200: %s",
+			rr.Body.String())
+	}
+}
+
+func TestPluginCallIsAdminOnly(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	// A real user-role key, not a bare one from config: sk-test/sk-user in the
+	// config both default to admin, so a test using them would pass the gate for
+	// the wrong reason and prove nothing.
+	rec, err := g.core.CreateKey("viewer", "user", nil, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	rr := doReqAs(t, g, rec.Key, http.MethodPost,
+		"/api/plugins/calltest/call/add", `{"x":1}`)
+	if rr.Code != http.StatusForbidden {
+		t.Fatalf("a non-admin key reached the call endpoint: HTTP %d %s",
+			rr.Code, rr.Body.String())
+	}
+}

internal/gateway/plugin_http_routes.go

@@ -0,0 +1,266 @@
+package gateway
+
+// Mount for plugin-registered HTTP endpoints.
+//
+// The gateway owns the listener, the authentication, the role check and the
+// response framing; the plugin owns the logic. That split is the point: a plugin
+// cannot open a socket, cannot bypass auth, and cannot choose a path outside
+// /api/plugins/http/.
+//
+// Why this exists rather than "plugins can call each other": some things are
+// worth querying from OUTSIDE the browser — a CI job asking whether the upstream
+// balance is above zero, a monitoring probe, another tool on the box. billing
+// never wanted this (it only receives hooks); a quota plugin does. Hard-coding a
+// route per plugin would mean a gateway change for every such plugin, and an
+// arbitrary plugin's code in the main mux is not something the routing table
+// should ever grow.
+
+import (
+	"net/http"
+	"sort"
+	"strconv"
+	"strings"
+
+	"llmsproxy/internal/lua"
+)
+
+// pluginHTTPPrefix is the mount point. It is deliberately narrow: a plugin's
+// routes live under one prefix, which makes them greppable, makes them easy to
+// revoke wholesale, and means a plugin cannot shadow /api/keys or /api/stats.
+const pluginHTTPPrefix = "/api/plugins/http/"
+
+// handlePluginHTTP dispatches a request to whichever plugin registered the
+// longest matching path.
+//
+// Longest-match rather than first-match: a plugin may register /health and
+// another /health/deep, and both must work. With registration order deciding it,
+// the outcome would depend on load order and change when a plugin is added.
+func (g *Gateway) handlePluginHTTP(w http.ResponseWriter, r *http.Request) {
+	rest := strings.TrimPrefix(r.URL.Path, pluginHTTPPrefix)
+	ps := g.core.Plugins()
+	if ps == nil {
+		writeError(w, http.StatusNotFound, "not_found", "no plugins loaded")
+		return
+	}
+
+	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]
+		p, ok := matchHandlerPath(h.Path, rest)
+		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)
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed",
+			"this endpoint accepts "+chosen.Method+" only")
+		return
+	}
+
+	role := reqRole(r.Context())
+	if !chosen.Allows(role, reqKeyName(r.Context())) {
+		writeError(w, http.StatusForbidden, "forbidden",
+			handlerAccessHint(*chosen))
+		return
+	}
+
+	body, err := readBody(r)
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+		return
+	}
+	req := &lua.HandlerRequest{
+		Method:  r.Method,
+		Path:    rest,
+		Query:   flattenQuery(r.URL.Query()),
+		Params:  params,
+		Headers: flattenHeader(r.Header),
+		Body:    body,
+		Role:    role,
+		Key:     reqKey(r.Context()),
+	}
+	resp, err := ps.InvokeHandler(handlerOwner(chosen), chosen.Handler, req)
+	if err != nil {
+		// A plugin error is a 502, not a 500: the gateway is fine, the plugin is
+		// not, and the operator needs to tell those apart in their logs.
+		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
+		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.
+func handlerAccessHint(h lua.HTTPHandler) string {
+	switch {
+	case len(h.Keys) > 0:
+		return "this endpoint allows only the named keys (" +
+			strings.Join(h.Keys, ", ") + "); authenticate with one of them"
+	case len(h.Roles) > 0:
+		return "this endpoint requires role " + strings.Join(h.Roles, " or ") +
+			" (widen it in the plugin's http declaration, or set public=true)"
+	default:
+		return "this endpoint declares no roles, so it is admin-only; " +
+			"set roles=… or keys=… in the plugin's http declaration"
+	}
+}
+
+// handlerOwner returns the plugin name that registered h. The gateway stores the
+// owner alongside the handler precisely so this does not have to be guessed from
+// the path prefix.
+func handlerOwner(h *lua.HTTPHandler) string { return h.Plugin }
+
+// segmentCount is how deep a route is, used for longest-match. A path with more
+// segments wins, so /health/deep beats /health when both are registered.
+func segmentCount(p string) int { return strings.Count(p, "/") }
+
+// matchHandlerPath matches a route pattern against a concrete path, capturing
+// ":name" segments.
+//
+// Returns ok=false on a segment-count mismatch or a missing value for a
+// param — a route must not match with an empty capture, or two different URLs
+// would reach the same handler indistinguishably.
+func matchHandlerPath(pattern, path string) (map[string]string, bool) {
+	pp := strings.Split(strings.Trim(pattern, "/"), "/")
+	sp := strings.Split(strings.Trim(path, "/"), "/")
+	if len(pp) != len(sp) {
+		return nil, false
+	}
+	var params map[string]string
+	for i := range pp {
+		if strings.HasPrefix(pp[i], ":") {
+			name := pp[i][1:]
+			if sp[i] == "" {
+				return nil, false
+			}
+			if params == nil {
+				params = map[string]string{}
+			}
+			params[name] = sp[i]
+			continue
+		}
+		if pp[i] != sp[i] {
+			return nil, false
+		}
+	}
+	return params, true
+}
+
+func flattenQuery(q map[string][]string) map[string]string {
+	out := make(map[string]string, len(q))
+	for k, v := range q {
+		if len(v) > 0 {
+			out[k] = v[0]
+		}
+	}
+	return out
+}
+
+func flattenHeader(h http.Header) map[string]string {
+	out := make(map[string]string, len(h))
+	for k := range h {
+		out[strings.ToLower(k)] = h.Get(k)
+	}
+	return out
+}
+
+func readBody(r *http.Request) (string, error) {
+	if r.Body == nil {
+		return "", nil
+	}
+	defer r.Body.Close()
+	// A cap, because the handler runs inside the gateway process: an unbounded
+	// body from an authenticated caller is a memory exhaustion vector.
+	const maxBody = 4 << 20
+	buf := make([]byte, 0, 4096)
+	tmp := make([]byte, 32*1024)
+	for {
+		n, err := r.Body.Read(tmp)
+		if n > 0 {
+			if len(buf)+n > maxBody {
+				return "", &bodyTooLarge{}
+			}
+			buf = append(buf, tmp[:n]...)
+		}
+		if err != nil {
+			break
+		}
+	}
+	return string(buf), nil
+}
+
+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

internal/gateway/plugins_api.go

@@ -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")
@@ -93,6 +144,60 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// /api/plugins/{name}/http/{path}/access — the operator's view of, and edit
+	// surface for, which keys an endpoint answers.
+	if i := strings.Index(path, "/http/"); i >= 0 {
+		name := path[:i]
+		rest := path[i+len("/http/"):]
+		if strings.HasSuffix(rest, "/access") {
+			epPath := strings.TrimSuffix(rest, "/access")
+			g.handleEndpointAccess(w, r, name, epPath)
+			return
+		}
+	}
+
+	// /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:
+	// channels.lua 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]
+		fn := path[i+len("/call/"):]
+		if name == "" || fn == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "call needs a plugin and a function name")
+			return
+		}
+		g.handlePluginCall(w, r, name, fn)
+		return
+	}
+
+	// /api/plugins/{name}/fanout/{fn} — one call per argument set, in
+	// parallel, from one request. See handlePluginFanout for why the
+	// request-per-item form cannot be made concurrent from the client.
+	if i := strings.Index(path, "/fanout/"); i >= 0 {
+		name := path[:i]
+		fn := path[i+len("/fanout/"):]
+		if name == "" || fn == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "fanout needs a plugin and a function name")
+			return
+		}
+		g.handlePluginFanout(w, r, name, fn)
+		return
+	}
+
 	if reqRole(r.Context()) != "admin" {
 		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
 		return
@@ -106,6 +211,11 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 				"on_disk":     ps.OnDisk(),
 				"hook_errors": ps.HookErrors(),
 				"plugin_dir":  g.pluginDir(),
+				// Visibility overrides and the endpoint access table, so the
+				// plugin page renders what is IN FORCE rather than making the
+				// operator read .lua files or config.yaml to find out.
+				"visibility":  ps.VisibilityOverrides(),
+				"http_access": ps.EndpointAccess(),
 			})
 		case http.MethodPost:
 			var body struct {
@@ -211,15 +321,33 @@ func (g *Gateway) handlePluginState(w http.ResponseWriter, r *http.Request, name
 	ps := g.core.Plugins()
 	switch r.Method {
 	case http.MethodGet:
-		// Any role may read: plugin state is reporting data (cost, counts),
-		// and the caller has already been authenticated. Admin-only would stop
-		// a user key's own billing widget from rendering.
+		// The plugin's audience gates its DATA as well as its page. Hiding the UI
+		// while leaving GET /state open would put the numbers one curl away, which
+		// is the leak the audience field exists to close — and billing's state is
+		// a global ledger, not the caller's own usage.
+		//
+		// A plugin that declares no audience is unchanged (every role may read
+		// it), so a single-purpose widget keeps working without edits.
+		role := reqRole(r.Context())
+		if !ps.StateVisibleTo(name, role) {
+			writeError(w, http.StatusForbidden, "forbidden",
+				"plugin "+name+" is not visible to role "+role)
+			return
+		}
 		// prices rides along: the billing rules editor must show the table
 		// that is actually in effect, not a reconstruction from config. It is
 		// absent for plugins with no prices (every plugin but billing).
+		state := ps.State(name)
+		// A plugin may SPLIT its state: global buckets stay deployment-wide while
+		// per-caller buckets are narrowed to the authenticated key. Role
+		// visibility alone cannot express this — the same role is allowed to read
+		// the plugin, just not another caller's row.
+		if scope := ps.StateScopeFor(name); scope != nil && role != "admin" {
+			state = splitStateForKey(state, scope, reqKey(r.Context()))
+		}
 		payload := map[string]interface{}{
 			"plugin": name,
-			"state":  ps.State(name),
+			"state":  state,
 		}
 		if pr := ps.Prices(name); pr != nil {
 			payload["prices"] = pr
@@ -245,6 +373,107 @@ func (g *Gateway) handlePluginState(w http.ResponseWriter, r *http.Request, name
 	}
 }
 
+// Call invokes a named function on a plugin's Lua state and returns its result.
+//
+// This exists because plugin pages otherwise have NO way to reach their own
+// plugin code: pluginAPI offers state read/write and a couple of callbacks, and
+// that is enough for billing (whose numbers arrive on hooks during request
+// forwarding) but useless for a plugin that must do work on demand. A quota
+// page has to ask the vendor when the operator clicks Refresh, and no hook
+// fires at that moment.
+//
+// It is deliberately narrow:
+//   - The function must be one the plugin explicitly lists in plugin.exports.
+//     Anything else is refused, so this endpoint cannot become a generic "run
+//     arbitrary Lua with my credentials" oracle over a plugin's private state.
+//   - The caller must be admin, and the result is JSON — no Lua value is ever
+//     handed back to Go as a live closure.
+//   - Errors come back as (nil, message) from the plugin, matching every other
+//     plugin entry point, and surface as a 502 rather than raising.
+func (g *Gateway) handlePluginCall(w http.ResponseWriter, r *http.Request, name, fn string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	if r.Method != http.MethodPost {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+		return
+	}
+	var args map[string]interface{}
+	if err := json.NewDecoder(r.Body).Decode(&args); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	result, err := g.core.Plugins().Call(name, fn, args)
+	if err != nil {
+		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{"result": result})
+}
+
+// handlePluginFanout runs one plugin export per element of a list, in parallel,
+// from a SINGLE request.
+//
+// Why this exists: the quota page has twenty cards, and it used to fire twenty
+// browser requests at /call/. They arrived concurrently and then serialised
+// inside Plugins.Call, which holds the plugin mutex across the Lua function —
+// including its network wait. Measured on the live gateway: 20 probes took 29s of
+// wall clock while a single probe with no contention took 1.4–3.3s, so the page
+// sat on "probing…" for half a minute. The browser-side concurrency was real and
+// made no difference.
+//
+// A fan-out cannot be assembled from the client either: any request-per-item
+// arrangement is exactly the shape that queues. The parallelism has to happen
+// where the fan-out is received.
+func (g *Gateway) handlePluginFanout(w http.ResponseWriter, r *http.Request, name, fn string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	if r.Method != http.MethodPost {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+		return
+	}
+	var req struct {
+		// Args is a LIST of argument objects, one per call. A bare object is
+		// rejected rather than coerced: "args" meaning "one call" and "args"
+		// meaning "many" depending on its shape is how a caller ends up probing
+		// one source and believing it probed twenty.
+		Args []map[string]interface{} `json:"args"`
+	}
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	if len(req.Args) == 0 {
+		writeError(w, http.StatusBadRequest, "invalid_request", "args must be a non-empty list of argument objects")
+		return
+	}
+	// The page legitimately has one card per configured source. The bound is not
+	// arbitrary: it stops one request from asking for thousands of VMs, and it is
+	// enforced before any work starts rather than discovered shard by shard.
+	if len(req.Args) > fanoutMaxArgs {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			fmt.Sprintf("args has %d entries, limit is %d", len(req.Args), fanoutMaxArgs))
+		return
+	}
+	results, err := g.core.Plugins().ShardCall(name, fn, req.Args)
+	if err != nil {
+		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
+		return
+	}
+	// A shard that failed contributes a null in its slot. That is deliberate and
+	// is NOT the same as the whole request failing: one unreachable upstream must
+	// not turn a page of twenty cards into a single error.
+	writeJSON(w, http.StatusOK, map[string]interface{}{"results": results})
+}
+
+// fanoutMaxArgs bounds one fan-out request. Twenty is the measured real source
+// count on the gateway this was built for; the headroom is for operators with
+// more sources configured.
+const fanoutMaxArgs = 64
+
 func (g *Gateway) pluginDir() string {
 	if c := g.core.Config(); c != nil {
 		return c.PluginDir
@@ -282,3 +511,241 @@ func (g *Gateway) removePlugin(name string) error {
 	_ = os.Remove(filepath.Join(g.pluginDir(), name+".lua"))
 	return g.core.Plugins().Unload(name)
 }
+
+// handleEndpointAccess serves the operator's view and edit surface for one
+// plugin endpoint's reachable keys.
+//
+//	GET    …/http/{path}/access   the declared rules plus any override
+//	PUT    …/http/{path}/access   body {"access": "admin,@homeagent"}
+//	DELETE …/http/{path}/access   clear the override
+//
+// The spec string is a comma list: a bare word is a role, "@name" is a gateway
+// key's configured name, and the literal "public" opens it to any
+// authenticated key. That is the same grammar the visibility override uses, and
+// keeping one grammar means an operator learns it once.
+//
+// Key names are validated against the configured keys BEFORE being stored. An
+// allowlist naming a key that does not exist is a silent no-op — every caller
+// is refused — and the operator would have no way to tell that apart from the
+// endpoint simply working.
+func (g *Gateway) handleEndpointAccess(w http.ResponseWriter, r *http.Request, name, epPath string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	ps := g.core.Plugins()
+
+	// epPath "*" is the WHOLE-PLUGIN visibility override, not an endpoint. It
+	// shares this route (and therefore this store) on purpose: an operator
+	// setting "who can see this plugin" and "who can call this endpoint" is
+	// asking the same question at two granularities, and two stores would let
+	// them disagree. SetEndpointAccess cannot serve it — it validates that the
+	// endpoint exists, which "*" by definition does not.
+	if epPath == pluginWholeOverride {
+		g.handlePluginVisibility(w, r, name)
+		return
+	}
+
+	switch r.Method {
+	case http.MethodGet:
+		for _, e := range ps.EndpointAccess() {
+			if e.Plugin == name && e.Path == epPath {
+				writeJSON(w, http.StatusOK, map[string]interface{}{"endpoint": e})
+				return
+			}
+		}
+		writeError(w, http.StatusNotFound, "not_found", "no such endpoint")
+	case http.MethodPut, http.MethodDelete:
+		spec := ""
+		if r.Method == http.MethodPut {
+			var body struct {
+				Access string `json:"access"`
+			}
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+				return
+			}
+			spec = body.Access
+		}
+		if err := ps.ValidateAccessSpec(spec); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		if err := ps.SetEndpointAccess(name, epPath, spec); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "access": spec})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+	}
+}
+
+// splitStateForKey rewrites a plugin state so that a non-admin caller sees only
+// the buckets that belong to them.
+//
+// The plugin declares which buckets are global and which are per-key. A per-key
+// bucket is keyed by the MASKED key ("***a1b2c3" — the same value the audit log
+// uses), so the caller's own row is found by masking the presented secret and
+// matching the last six characters.
+//
+// Keys that do not match are DROPPED, not zeroed: a zeroed row would render as
+// "you spent nothing" and, worse, would leak that a row exists for someone else.
+func splitStateForKey(state interface{}, scope *lua.StateScope, key string) interface{} {
+	m, ok := state.(map[string]interface{})
+	if !ok || scope == nil || key == "" {
+		return state
+	}
+	masked := maskKey(key)
+	out := make(map[string]interface{}, len(m))
+	for k, v := range m {
+		if !scope.IsGlobal(k) {
+			// Per-key bucket: keep only the caller's rows, at every nesting level
+			// the scope mentions (by_day_key is {day: {masked: bucket}}).
+			out[k] = filterKeyBuckets(v, masked, scope.DepthOf(k))
+			continue
+		}
+		out[k] = v
+	}
+	return out
+}
+
+// filterKeyBuckets narrows a keyed structure to one masked key.
+//
+// depth 1 is {masked: bucket}; depth 2 is {day: {masked: bucket}}. The level is
+// declared rather than sniffed, because both shapes are dicts and sniffing would
+// confuse an empty bucket with a nested one.
+func filterKeyBuckets(v interface{}, masked string, depth int) interface{} {
+	switch depth {
+	case 0:
+		return v // not a per-key bucket; leave alone
+	case 1:
+		m, ok := v.(map[string]interface{})
+		if !ok {
+			return v
+		}
+		if row, ok := m[masked]; ok {
+			return map[string]interface{}{masked: row}
+		}
+		return map[string]interface{}{}
+	default:
+		m, ok := v.(map[string]interface{})
+		if !ok {
+			return v
+		}
+		out := make(map[string]interface{}, len(m))
+		for day, inner := range m {
+			im, ok := inner.(map[string]interface{})
+			if !ok {
+				continue
+			}
+			if row, ok := im[masked]; ok {
+				out[day] = map[string]interface{}{masked: row}
+			}
+		}
+		return out
+	}
+}
+
+// pluginWholeOverride is the epPath that addresses a plugin's overall visibility
+// rather than one of its endpoints.
+const pluginWholeOverride = "*"
+
+// handlePluginVisibility is the whole-plugin half of the access API. The
+// audience string uses the same vocabulary as an endpoint spec (roles, @key
+// names, public) so the UI can drive both from one input.
+func (g *Gateway) handlePluginVisibility(w http.ResponseWriter, r *http.Request, name string) {
+	ps := g.core.Plugins()
+	switch r.Method {
+	case http.MethodGet:
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"plugin":     name,
+			"visibility": ps.VisibilityOverrides()[name],
+		})
+	case http.MethodPut, http.MethodDelete:
+		spec := ""
+		if r.Method == http.MethodPut {
+			var body struct {
+				Access string `json:"access"`
+			}
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+				return
+			}
+			spec = body.Access
+		}
+		if err := ps.ValidateAccessSpec(spec); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		ps.SetPluginVisibility(name, spec)
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "visibility": spec})
+	default:
+		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

@@ -210,3 +210,54 @@ func TestPluginStatePUTIsAdminOnly(t *testing.T) {
 		t.Errorf("admin PUT state = %d: %s", adm.Code, adm.Body.String())
 	}
 }
+
+// TestPluginsListCarriesAccessTables: the plugin page renders visibility
+// overrides and the endpoint access table from /api/plugins. Without them in the
+// response the operator has to open a .lua or config.yaml to learn who can see
+// what, which is exactly what the UI is meant to avoid.
+func TestPluginsListCarriesAccessTables(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodGet, "/api/plugins", "")
+	if rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/plugins = %d", rr.Code)
+	}
+	var out struct {
+		Visibility map[string]string        `json:"visibility"`
+		HTTPAccess []map[string]interface{} `json:"http_access"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode: %v (body=%s)", err, rr.Body.String())
+	}
+	// The calltest plugin registers no endpoints, so the table is present but
+	// empty — the KEY is that the field exists at all. A missing field decodes to
+	// nil and the UI silently renders nothing, which looks identical to "this
+	// plugin has no endpoints".
+	if !strings.Contains(rr.Body.String(), "http_access") {
+		t.Error("/api/plugins response lacks the http_access field")
+	}
+	if !strings.Contains(rr.Body.String(), "visibility") {
+		t.Error("/api/plugins response lacks the visibility field")
+	}
+}
+
+// TestPluginWholeVisibilityViaAccessAPI covers the "*" path: the plugin page
+// edits a plugin's overall audience through the same route it edits endpoints
+// with. Before this existed the UI had a visibility selector whose save 404'd,
+// because SetEndpointAccess validates that the endpoint exists and "*" does not.
+func TestPluginWholeVisibilityViaAccessAPI(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodPut, "/api/plugins/calltest/http/*/access",
+		`{"access":"admin, user"}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("PUT */access = %d: %s", rr.Code, rr.Body.String())
+	}
+	if !strings.Contains(rr.Body.String(), `"ok":true`) {
+		t.Errorf("unexpected body: %s", rr.Body.String())
+	}
+	// A user role is a legitimate audience; a nonsense one is not.
+	bad := doReq(t, g, http.MethodPut, "/api/plugins/calltest/http/*/access",
+		`{"access":"../../etc"}`)
+	if bad.Code != http.StatusBadRequest {
+		t.Errorf("traversal-shaped audience = %d, want 400: %s", bad.Code, bad.Body.String())
+	}
+}

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"
+
+	"llmsproxy/internal/lua"
+	"llmsproxy/internal/provider"
+	"llmsproxy/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"
+
+	"llmsproxy/internal/lua"
+	"llmsproxy/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

@@ -216,6 +216,14 @@ 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":
@@ -224,6 +232,11 @@ 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/"):
@@ -235,6 +248,16 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 		// 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":
@@ -305,20 +328,43 @@ func (g *Gateway) auth(next http.Handler) http.Handler {
 			http.Redirect(w, r, "/login?continue="+url.QueryEscape(r.URL.Path), http.StatusFound)
 			return
 		}
-		next.ServeHTTP(w, r.WithContext(withAuth(r.Context(), key, rec.Role)))
+		next.ServeHTTP(w, r.WithContext(withAuthNamed(r.Context(), key, rec.Role, rec.Name)))
 	})
 }
 
-// authCtx carries the authenticated gateway key and its role.
+// authCtx carries the authenticated gateway key, its role, and the key's
+// configured NAME.
+//
+// The name is here because plugin HTTP endpoints authorise by key name rather
+// than by secret, and looking the name up from the key at the point of use would
+// mean reading config on every request. It is "" for a key configured without
+// one, which such an endpoint can never match.
 type authCtx struct {
 	key  string
 	role string
+	name string
 }
 
 type authCtxKeyT struct{}
 
 func withAuth(ctx context.Context, key, role string) context.Context {
-	return context.WithValue(ctx, authCtxKeyT{}, authCtx{key: key, role: role})
+	return withAuthNamed(ctx, key, role, "")
+}
+
+// withAuthNamed is withAuth plus the key's configured name.
+func withAuthNamed(ctx context.Context, key, role, name string) context.Context {
+	return context.WithValue(ctx, authCtxKeyT{}, authCtx{key: key, role: role, name: name})
+}
+
+// reqKeyName returns the authenticated key's configured name, or "" if it has
+// none. Plugin endpoint authorisation is written against this, not against the
+// secret: a secret is rotated without notice, so a rule keyed on it would
+// silently stop matching.
+func reqKeyName(ctx context.Context) string {
+	if a, ok := ctx.Value(authCtxKeyT{}).(authCtx); ok {
+		return a.name
+	}
+	return ""
 }
 
 // reqKey returns the authenticated gateway key id (masked suffix for display).

internal/gateway/stats.go

@@ -43,7 +43,17 @@ type Req struct {
 	// accounting at all (even when the hit count is 0). The WebUI shows
 	// "0%" instead of "—" for such rows.
 	CacheReported bool `json:"cache_reported,omitempty"`
-	OK            bool `json:"ok"`
+	// Reasoning is completion_tokens_details.reasoning_tokens.
+	//
+	// It is a SUBSET of Compl, not an addition: reasoning tokens are billed at
+	// the completion rate and already counted there. This field exists so a
+	// pricing rule can split them when a provider prices thinking tokens
+	// differently (most do NOT — OpenAI and DeepSeek bill them as ordinary
+	// output), and so a reader can see what fraction of a bill is thinking.
+	// 0 means "not reported", which is NOT the same as "no reasoning happened" —
+	// three of the sources measured return no such field at all.
+	Reasoning int64 `json:"reasoning_tokens,omitempty"`
+	OK        bool  `json:"ok"`
 	// Status http status code
 	Status int `json:"status"`
 	// Err short error message
@@ -755,9 +765,20 @@ func (s *Stats) WindowTokens(model, source string, sec int64) int64 {
 	return total
 }
 
+// rows flattens a map[string]*Stat into a reportable, descending-by-requests
+// list.
+//
+// A nil value is skipped rather than dereferenced. The maps this is called with
+// are built by several callers (Snapshot, PeriodSnapshot, the audit replay) and a
+// nil entry is a legitimate representation of "this dimension has no data" —
+// `map[string]*Stat{key: s.byKey[key]}` produced exactly that whenever the
+// queried key had no recorded traffic, and `*v` then panicked the handler.
 func rows(m map[string]*Stat) []StatsRow {
 	out := make([]StatsRow, 0, len(m))
 	for k, v := range m {
+		if v == nil {
+			continue
+		}
 		out = append(out, StatsRow{Name: k, Stat: *v})
 	}
 	for i := 1; i < len(out); i++ {
@@ -1175,7 +1196,26 @@ func (s *Stats) Snapshot(limit int, key string) map[string]interface{} {
 	if key == "" {
 		byKey, byModel, bySrc = s.byKey, s.byModel, s.bySrc
 	} else {
-		byKey = map[string]*Stat{key: s.byKey[key]}
+		// The map is keyed by the MASKED key id (keyID(req.Key)), which is what
+		// by_key reports and what the dashboard sends as ?key=. An admin may
+		// pass anything in that query parameter, including the full key, and
+		// `s.byKey[key]` then returns the zero value — a nil *Stat.
+		//
+		// rows() below does `*v` on every value, so a nil entry panicked the
+		// handler: any ?key=<full key> request killed the response with an empty
+		// reply, and logged a full panic stack. Reproduced against the live
+		// gateway: /api/stats?period=day&limit=1&key=sk-gw-... → curl exit 52.
+		//
+		// A key with no recorded traffic is a legitimate question ("what did this
+		// key use?"), and its answer is zero — so an absent entry is an empty
+		// result, not an error, and not a crash.
+		st := s.byKey[key]
+		byKey = make(map[string]*Stat, 1)
+		if st != nil {
+			byKey[key] = st
+		} else {
+			byKey[key] = &Stat{}
+		}
 		byModel = s.byKeyModel[key]
 		bySrc = s.byKeySrc[key]
 	}
@@ -1192,8 +1232,11 @@ func (s *Stats) Snapshot(limit int, key string) map[string]interface{} {
 		}
 	}
 	bs := make([]agrRow, 0, len(s.byStatus))
-	for code := range s.byStatus {
-		bs = append(bs, agrRow{Name: strconv.Itoa(code), Stat: *s.byStatus[code]})
+	for code, st := range s.byStatus {
+		if st == nil {
+			continue
+		}
+		bs = append(bs, agrRow{Name: strconv.Itoa(code), Stat: *st})
 	}
 	sort.Slice(bs, func(i, j int) bool {
 		ci, _ := strconv.Atoi(bs[i].Name)
@@ -1211,3 +1254,16 @@ func (s *Stats) Snapshot(limit int, key string) map[string]interface{} {
 		"replay_partial": s.replayPartial,
 	}
 }
+
+// Active reports requests in flight right now.
+//
+// It is a live gauge, not an aggregate: there is no calendar period that scopes
+// it, which is why the period-scoped /api/stats response has to carry it
+// explicitly. Before it did, the day/week/month views fed `st.active || 0` a
+// missing key and showed a hard 0 — the same gateway that was serving hundreds
+// of requests a second looked idle in three of the four views.
+func (s *Stats) Active() int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.active
+}

internal/gateway/stats_period.go

@@ -174,6 +174,9 @@ func (s *Stats) PeriodSnapshot(p Period, key string, now time.Time) PeriodSnapsh
 	out.Srcs = rows(bySrc)
 	bs := make([]agrRow, 0, len(byStatus))
 	for code, st := range byStatus {
+		if st == nil {
+			continue
+		}
 		bs = append(bs, agrRow{Name: code, Stat: *st})
 	}
 	sort.Slice(bs, func(i, j int) bool {
@@ -192,7 +195,11 @@ func (s *Stats) PeriodSnapshot(p Period, key string, now time.Time) PeriodSnapsh
 	sort.Strings(ks)
 	out.Bucket = make([]PeriodBucket, 0, len(ks))
 	for _, k := range ks {
-		out.Bucket = append(out.Bucket, PeriodBucket{Bucket: k, Stat: *buckets[k]})
+		// Defensive for the same reason rows() is: ks comes from the bucket map's
+		// keys, and a nil value would panic on the deref below.
+		if st := buckets[k]; st != nil {
+			out.Bucket = append(out.Bucket, PeriodBucket{Bucket: k, Stat: *st})
+		}
 	}
 
 	// The window is only "complete" if the audit walk actually reached back

internal/gateway/stats_period_records_test.go

@@ -0,0 +1,159 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/url"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+// The dashboard paints its records table from the `records` field of
+// /api/stats (paintRecords(st.records)). The period branch of that handler
+// used to hand-write a response map and simply omit the key, so switching the
+// selector to day / week / month rendered an EMPTY table no matter how much
+// traffic the window had — while the lifetime view worked, because it goes
+// through Snapshot() which carries records.
+//
+// These tests pin the contract across every period: same records, same cursor.
+func TestStatsPeriodViewsShipRecords(t *testing.T) {
+	td := t.TempDir()
+	g := newTestGateway(t)
+
+	// Write audit records the period aggregator and the record pager both read.
+	g.stats.auditPath = filepath.Join(td, "audit.jsonl")
+	now := time.Now().UnixMilli()
+	for i := 0; i < 5; i++ {
+		g.stats.Record(Req{
+			Time: now - int64(i)*1000, Type: "chat", OK: true,
+			Model: "m1", Source: "s1", Status: 200,
+			Prompt: 10, Compl: 2,
+		})
+	}
+
+	for _, p := range []string{"day", "week", "month", "all"} {
+		rr := doReq(t, g, "GET", "/api/stats?period="+p, "")
+		if rr.Code != 200 {
+			t.Fatalf("period=%s: HTTP %d: %s", p, rr.Code, rr.Body.String())
+		}
+		var st map[string]interface{}
+		if err := json.Unmarshal(rr.Body.Bytes(), &st); err != nil {
+			t.Fatalf("period=%s: %v", p, err)
+		}
+		recs, ok := st["records"].([]interface{})
+		if !ok {
+			t.Fatalf("period=%s: response has no `records` array — the dashboard "+
+				"records table renders empty for this view (keys: %v)", p, keysOf(st))
+		}
+		if len(recs) != 5 {
+			t.Fatalf("period=%s: got %d records, want 5", p, len(recs))
+		}
+		// The scroll handler pages from next_cursor; without it the table is
+		// stuck at one screen no matter how far the user scrolls.
+		if _, ok := st["next_cursor"]; !ok {
+			t.Fatalf("period=%s: no next_cursor — paging can never continue", p)
+		}
+	}
+}
+
+// The records shipped with a period view must be the newest-first page the
+// paging endpoint would return, in the same order. If /api/stats and
+// /api/stats/records disagreed, scrolling would duplicate or skip rows.
+func TestStatsPeriodRecordsMatchPager(t *testing.T) {
+	td := t.TempDir()
+	g := newTestGateway(t)
+	g.stats.auditPath = filepath.Join(td, "audit.jsonl")
+	now := time.Now().UnixMilli()
+	for i := 0; i < 8; i++ {
+		g.stats.Record(Req{
+			Time: now - int64(i)*1000, Type: "chat", OK: true,
+			Model: "m1", Status: 200, Prompt: 5, Compl: 1,
+		})
+	}
+
+	rr := doReq(t, g, "GET", "/api/stats?period=day&limit=100", "")
+	var st struct {
+		Records []struct {
+			Time  int64  `json:"time"`
+			Model string `json:"model"`
+		} `json:"records"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &st); err != nil {
+		t.Fatal(err)
+	}
+	rr2 := doReq(t, g, "GET", "/api/stats/records?limit=100", "")
+	var pg struct {
+		Records []struct {
+			Time  int64  `json:"time"`
+			Model string `json:"model"`
+		} `json:"records"`
+	}
+	if err := json.Unmarshal(rr2.Body.Bytes(), &pg); err != nil {
+		t.Fatal(err)
+	}
+	if len(st.Records) != len(pg.Records) {
+		t.Fatalf("period view shipped %d records, pager returned %d — the two "+
+			"sources must agree or scrolling duplicates/skips rows",
+			len(st.Records), len(pg.Records))
+	}
+	for i := range st.Records {
+		if st.Records[i].Time != pg.Records[i].Time {
+			t.Fatalf("record %d differs: period view t=%d, pager t=%d",
+				i, st.Records[i].Time, pg.Records[i].Time)
+		}
+	}
+	// The two sources must agree on DIRECTION as well as content. Which
+	// direction that is depends on how many reverse-read chunks the audit file
+	// spans (a small file is read whole, a large one chunk-by-chunk), so the
+	// test asserts agreement rather than hard-coding newest-first — but they
+	// can never disagree, because the table interleaves both streams.
+	if len(st.Records) > 1 {
+		periodAsc := st.Records[1].Time > st.Records[0].Time
+		pagerAsc := pg.Records[1].Time > pg.Records[0].Time
+		if periodAsc != pagerAsc {
+			t.Fatalf("direction differs: period view %s, pager %s — the table "+
+				"would splice the two streams out of order",
+				map[bool]string{true: "oldest-first", false: "newest-first"}[periodAsc],
+				map[bool]string{true: "oldest-first", false: "newest-first"}[pagerAsc])
+		}
+	}
+}
+
+// A non-admin caller cannot widen their view by passing someone else's key in
+// the query string: exportKey overrides ?key= with the caller's own masked id.
+// This is a data-leak boundary, and it applies to the records the period view
+// ships just as much as to the aggregates.
+func TestStatsPeriodRecordsIgnoreForeignKeyForUsers(t *testing.T) {
+	td := t.TempDir()
+	g := newTestGateway(t)
+	g.stats.auditPath = filepath.Join(td, "audit.jsonl")
+
+	other := keyID("sk-gw-cccccccccccccccc")
+	own := keyID("sk-test")
+	now := time.Now().UnixMilli()
+	g.stats.Record(Req{Time: now, OK: true, Model: "m", Status: 200, Key: other})
+	g.stats.Record(Req{Time: now - 1000, OK: true, Model: "m", Status: 200, Key: own})
+
+	// sk-test authenticates as a non-admin (newTestGateway seeds it under
+	// gateway_keys, not keys), so ?key=<other> must be ignored and the caller
+	// sees only its own rows.
+	rr := doReq(t, g, "GET",
+		"/api/stats?period=day&key="+url.QueryEscape(other), "")
+	if rr.Code != 200 {
+		t.Fatalf("HTTP %d: %s", rr.Code, rr.Body.String())
+	}
+	var st struct {
+		Records []struct {
+			Key string `json:"key"`
+		} `json:"records"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &st); err != nil {
+		t.Fatal(err)
+	}
+	for _, r := range st.Records {
+		if r.Key == other {
+			t.Fatalf("another key's record leaked into a user view "+
+				"(caller=%s, requested=%s)", own, other)
+		}
+	}
+}

internal/gateway/stats_period_test.go

@@ -319,3 +319,116 @@ func TestPeriodFromQuery(t *testing.T) {
 		}
 	}
 }
+
+// TestPeriodWeekMayStartBeforeMonth pins a fact that looks like a bug and
+// gets reported as one: the calendar periods are NOT nested.
+//
+// On 2026-10-03 (a Saturday) the ISO week runs Mon 2026-09-28 .. Sat
+// 10-03, while the month window is 10-01 .. 10-03. The week therefore reaches
+// back into the previous month and its total is legitimately LARGER than the
+// month's. Anyone reading only the totals will conclude the aggregator is
+// broken; it is not, and this test exists so a future "fix" that makes the
+// windows nested (e.g. clamping the week to the month) is caught as the
+// regression it would be.
+func TestPeriodWeekMayStartBeforeMonth(t *testing.T) {
+	// Saturday 2026-10-03, UTC.
+	now := time.Date(2026, 10, 3, 12, 0, 0, 0, time.UTC)
+
+	weekFrom := periodStart(PeriodWeek, now)
+	monthFrom := periodStart(PeriodMonth, now)
+
+	weekFromD := time.UnixMilli(weekFrom).UTC()
+	monthFromD := time.UnixMilli(monthFrom).UTC()
+
+	if got := weekFromD.Format("2006-01-02"); got != "2026-09-28" {
+		t.Fatalf("week window start = %s, want 2026-09-28 (the Monday)", got)
+	}
+	if got := monthFromD.Format("2006-01-02"); got != "2026-10-01" {
+		t.Fatalf("month window start = %s, want 2026-10-01", got)
+	}
+	if !weekFromD.Before(monthFromD) {
+		t.Fatal("this test only means anything when the week starts before the month")
+	}
+	// The week must reach strictly further back than the month window does,
+	// otherwise the nesting intuition would hold and the UI note would be noise.
+	if back := monthFromD.Sub(weekFromD).Hours() / 24; back <= 0 {
+		t.Fatalf("week should start at least a day before the month window, got %v days back", back)
+	}
+}
+
+// The same overlap must be visible in the aggregate numbers, not just in the
+// window bounds: a week total larger than a month total is the exact symptom
+// an operator reported.
+func TestPeriodWeekTotalCanExceedMonthTotal(t *testing.T) {
+	dir := t.TempDir()
+	audit := filepath.Join(dir, "audit.jsonl")
+
+	// Two records before the month window (inside this ISO week) and one
+	// inside October.
+	pre := time.Date(2026, 9, 29, 10, 0, 0, 0, time.UTC).UnixMilli()
+	in1 := time.Date(2026, 10, 1, 10, 0, 0, 0, time.UTC).UnixMilli()
+	in2 := time.Date(2026, 10, 3, 10, 0, 0, 0, time.UTC).UnixMilli()
+
+	auditPeriod(t, audit,
+		Req{Time: pre, OK: true, Prompt: 100, Compl: 10},
+		Req{Time: pre + 1000, OK: true, Prompt: 100, Compl: 10},
+		Req{Time: in1, OK: true, Prompt: 50, Compl: 5},
+		Req{Time: in2, OK: true, Prompt: 50, Compl: 5},
+	)
+
+	s := periodStats(t, audit)
+	now := time.Date(2026, 10, 3, 12, 0, 0, 0, time.UTC)
+
+	week := s.PeriodSnapshot(PeriodWeek, "", now)
+	month := s.PeriodSnapshot(PeriodMonth, "", now)
+
+	if week.Total.Reqs != 4 {
+		t.Fatalf("week reqs = %d, want 4 (both September records fall in this ISO week)", week.Total.Reqs)
+	}
+	if month.Total.Reqs != 2 {
+		t.Fatalf("month reqs = %d, want 2", month.Total.Reqs)
+	}
+	if week.Total.Tokens <= month.Total.Tokens {
+		t.Fatalf("week tokens (%d) must exceed month tokens (%d): the week reaches into September",
+			week.Total.Tokens, month.Total.Tokens)
+	}
+	// Day must be a strict subset of both, which is the only nesting that holds.
+	day := s.PeriodSnapshot(PeriodDay, "", now)
+	if day.Total.Reqs != 1 {
+		t.Fatalf("day reqs = %d, want 1", day.Total.Reqs)
+	}
+	if day.Total.Tokens >= month.Total.Tokens || day.Total.Tokens >= week.Total.Tokens {
+		t.Fatal("today must be contained in both the week and the month totals")
+	}
+}
+
+// TestPeriodStatsCarryActiveGauge pins the live "active requests" gauge onto
+// every period response.
+//
+// It is a live gauge, not an aggregate, so no calendar window can scope it — but
+// the period response is what the day/week/month views render from, and it used
+// to omit the field entirely. The UI then evaluated `st.active || 0` and printed
+// a hard 0 in three of four views, while the all-time view (whose response did
+// carry the field) showed the truth. Reading "0 active" off a busy gateway is
+// worse than reading nothing: it is a confident wrong answer.
+func TestPeriodStatsCarryActiveGauge(t *testing.T) {
+	for _, period := range []string{"day", "week", "month", "all"} {
+		g := newTestGateway(t)
+		rr := doReq(t, g, http.MethodGet, "/api/stats?period="+period, "")
+		if rr.Code != http.StatusOK {
+			t.Fatalf("period=%s: %d", period, rr.Code)
+		}
+		var out map[string]interface{}
+		if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+			t.Fatalf("period=%s decode: %v", period, err)
+		}
+		v, ok := out["active"]
+		if !ok {
+			t.Errorf("period=%s response has no active field; the UI renders 0", period)
+			continue
+		}
+		if _, ok := v.(float64); !ok {
+			t.Errorf("period=%s active is %T, want a number", period, v)
+		}
+	}
+}

internal/gateway/stats_unknown_key_test.go

@@ -0,0 +1,140 @@
+package gateway
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+)
+
+// TestSnapshotUnknownKeyDoesNotPanic covers a nil-dereference that killed the
+// response for any /api/stats request carrying a ?key= the statistics had no
+// entry for.
+//
+// The maps in Snapshot are keyed by the MASKED key id (keyID(req.Key)) — that
+// is what by_key reports and what the dashboard sends. An admin may pass anything
+// in the query parameter, including the full key, and `s.byKey[key]` then yields
+// the zero value: a nil *Stat. rows() dereferenced every value, so the handler
+// panicked and the client got an empty reply plus a panic stack in the log.
+//
+// Reproduced against the live gateway before the fix:
+//
+//	/api/stats?period=day&limit=1&key=sk-gw-7ffe12c5cbacec40bd39a7291df9a9ed
+//	→ curl: (52) Empty reply from server
+//	→ log: http: panic serving …: invalid memory address or nil pointer dereference
+//	    llmsproxy/internal/gateway.rows  stats.go:771
+//
+// A key with no recorded traffic is a legitimate question, and zero is its
+// answer — not a crash.
+func TestSnapshotUnknownKeyDoesNotPanic(t *testing.T) {
+	s := NewStats(16)
+	// One recorded key, masked the way Record masks it.
+	s.Record(Req{Key: "***aaaaaa", Type: "chat", Model: "m", Prompt: 5, Compl: 2})
+
+	cases := []struct {
+		name string
+		key  string
+	}{
+		{"masked key that exists", "***aaaaaa"},
+		{"masked key that does not exist", "***zzzzzz"},
+		{"full key rather than the mask", "sk-gw-full-key-value"},
+		{"arbitrary text", "not-a-key-at-all"},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			// A panic here fails the subtest by unwinding through the test
+			// goroutine; rows() used to take the whole process with it.
+			out := s.Snapshot(10, c.key)
+			byKey, ok := out["by_key"].([]StatsRow)
+			if !ok {
+				t.Fatalf("by_key is %T, want []StatsRow", out["by_key"])
+			}
+			total, _ := out["total"].(Stat)
+			t.Logf("key=%q → by_key %d 条, total.reqs=%d", c.key, len(byKey), total.Reqs)
+		})
+	}
+}
+
+// TestSnapshotUnknownKeyReportsZeroNotAbsence pins the semantics: a key with no
+// traffic answers zero, and the by_key row still exists so a dashboard that keys
+// off its presence renders a row of zeros rather than dropping the key.
+func TestSnapshotUnknownKeyReportsZeroNotAbsence(t *testing.T) {
+	s := NewStats(16)
+	out := s.Snapshot(10, "***zzzzzz")
+	byKey, _ := out["by_key"].([]StatsRow)
+	if len(byKey) != 1 {
+		t.Fatalf("by_key has %d rows, want 1: the queried key must appear with zeros", len(byKey))
+	}
+	if byKey[0].Name != "***zzzzzz" {
+		t.Errorf("row name = %q, want ***zzzzzz", byKey[0].Name)
+	}
+	if byKey[0].Reqs != 0 {
+		t.Errorf("reqs = %d, want 0", byKey[0].Reqs)
+	}
+}
+
+// TestPeriodSnapshotUnknownKeyDoesNotPanic is the same guard on the period view,
+// which builds its own buckets and calls the same flattening helpers.
+func TestPeriodSnapshotUnknownKeyDoesNotPanic(t *testing.T) {
+	dir := t.TempDir()
+	audit := dir + "/audit.jsonl"
+	writeAuditFile(t, audit, 3, "***aaaaaa", time.Now().Add(-time.Hour).UnixMilli())
+	s := periodStats(t, audit)
+
+	for _, key := range []string{"***aaaaaa", "***zzzzzz", "sk-gw-full-key-value"} {
+		out := s.PeriodSnapshot(PeriodAll, key, time.Now())
+		t.Logf("key=%q → total.reqs=%d byKey=%d buckets=%d",
+			key, out.Total.Reqs, len(out.ByKey), len(out.Bucket))
+	}
+}
+
+// TestRowsSkipsNilValues pins the defensive behaviour directly, so a future
+// caller that reintroduces a nil entry fails here rather than in production.
+func TestRowsSkipsNilValues(t *testing.T) {
+	in := map[string]*Stat{
+		"has-traffic": {Reqs: 5},
+		"no-entry":    nil,
+	}
+	got := rows(in)
+	if len(got) != 1 {
+		t.Fatalf("rows() returned %d rows, want 1 (the nil entry must be skipped): %+v", len(got), got)
+	}
+	if got[0].Name != "has-traffic" {
+		t.Errorf("kept %q, want has-traffic", got[0].Name)
+	}
+}
+
+// TestRowsEmptyIsEmptyNotNil keeps the JSON shape honest: an all-nil map yields
+// an empty list, which encodes as [] rather than null — the frontend does
+// Array.isArray checks on these fields.
+func TestRowsEmptyIsEmptyNotNil(t *testing.T) {
+	got := rows(map[string]*Stat{"a": nil})
+	if got == nil {
+		t.Fatal("rows() returned nil; the report field must stay [] instead of null")
+	}
+	if len(got) != 0 {
+		t.Fatalf("got %d rows, want 0", len(got))
+	}
+}
+
+// TestStatsAPIUnknownKeyReturnsJSON exercises the HTTP path so the assertion
+// covers the reported symptom (an empty reply), not just the helper.
+func TestStatsAPIUnknownKeyReturnsJSON(t *testing.T) {
+	g := newTestGateway(t)
+	srv := httptest.NewServer(g.Handler())
+	defer srv.Close()
+
+	req, err := http.NewRequest("GET", srv.URL+"/api/stats?period=day&limit=1&key=sk-gw-full-key", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header.Set("Authorization", "Bearer sk-test")
+	resp, err := srv.Client().Do(req)
+	if err != nil {
+		t.Fatalf("request failed (an empty reply is what the live gateway produced): %v", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, want 200", resp.StatusCode)
+	}
+}

internal/gateway/thinking_test.go

@@ -0,0 +1,133 @@
+package gateway
+
+import (
+	"encoding/json"
+	"testing"
+
+	"llmsproxy/internal/types"
+)
+
+// TestThinkingFromClient 把「客户端措辞 → 中立语义」这条翻译链钉死。
+//
+// v1 的问题：disable_thinking 是内核私有字段，被序列化进交给适配器的 body，
+// 于是 10 个适配器各自写 req.disable_thinking = nil 来清理它。v2 把它换成
+// 双方都认可的 thinking 字段，这条测试确保翻译只发生一次、且在适配器之前。
+func TestThinkingFromClient(t *testing.T) {
+	cases := []struct {
+		name string
+		req  chatRequest
+		want *types.ThinkingConfig
+	}{
+		{
+			name: "无任何字段 → 不带 thinking（上游用自己默认）",
+			req:  chatRequest{},
+			want: nil,
+		},
+		{
+			name: "disable_thinking=true → enabled=false",
+			req:  chatRequest{DisableThinking: true},
+			want: &types.ThinkingConfig{Enabled: boolPtr(false)},
+		},
+		{
+			name: "中立 thinking 显式给出 → 原样透传",
+			req:  chatRequest{Thinking: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 4096}},
+			want: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 4096},
+		},
+		{
+			name: "两者冲突时，中立的 thinking 优先（它带预算，布尔表达不了）",
+			req: chatRequest{
+				DisableThinking: true,
+				Thinking:        &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 1024},
+			},
+			want: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 1024},
+		},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := c.req.thinkingFromClient()
+			if c.want == nil {
+				if got != nil {
+					t.Fatalf("want nil, got %+v", got)
+				}
+				return
+			}
+			if got == nil {
+				t.Fatal("want non-nil thinking, got nil")
+			}
+			if (got.Enabled == nil) != (c.want.Enabled == nil) {
+				t.Fatalf("Enabled presence mismatch: got %v want %v", got.Enabled, c.want.Enabled)
+			}
+			if got.Enabled != nil && *got.Enabled != *c.want.Enabled {
+				t.Errorf("Enabled = %v, want %v", *got.Enabled, *c.want.Enabled)
+			}
+			if got.BudgetTokens != c.want.BudgetTokens {
+				t.Errorf("BudgetTokens = %d, want %d", got.BudgetTokens, c.want.BudgetTokens)
+			}
+		})
+	}
+}
+
+// TestDisableThinkingNeverReachesAdapterBody 是那条泄漏的回归测试：内核私有
+// 字段不得出现在交给适配器的 body 里。
+func TestDisableThinkingNeverReachesAdapterBody(t *testing.T) {
+	var r chatRequest
+	if err := json.Unmarshal([]byte(`{"model":"m","messages":[],"disable_thinking":true}`), &r); err != nil {
+		t.Fatal(err)
+	}
+	inner := &types.ChatRequest{
+		Model:         "m",
+		Thinking:      r.thinkingFromClient(),
+		Tools:         r.rawTools(),
+		ToolChoice:    r.rawToolChoice(),
+		ExtraBody:     r.ExtraBody,
+		ClientSession: "",
+	}
+	body, err := json.Marshal(inner)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]json.RawMessage
+	if err := json.Unmarshal(body, &m); err != nil {
+		t.Fatal(err)
+	}
+	if _, leaked := m["disable_thinking"]; leaked {
+		t.Errorf("内核私有字段 disable_thinking 泄漏进适配器 body（v1 回归）: %s", body)
+	}
+	if _, ok := m["thinking"]; !ok {
+		t.Errorf("中立 thinking 字段缺失: %s", body)
+	}
+}
+
+// TestToolsPassthroughRaw 确认 tools/tool_choice 以原始形状交给适配器：
+// 内核不解释工具 schema，那是适配器的职责。
+func TestToolsPassthroughRaw(t *testing.T) {
+	var r chatRequest
+	body := `{"model":"m","messages":[],"tools":[{"type":"function","function":{"name":"f","parameters":{"type":"object"}}}],"tool_choice":"auto"}`
+	if err := json.Unmarshal([]byte(body), &r); err != nil {
+		t.Fatal(err)
+	}
+	tools := r.rawTools()
+	if len(tools) == 0 {
+		t.Fatal("tools 未透传")
+	}
+	var decoded []map[string]interface{}
+	if err := json.Unmarshal(tools, &decoded); err != nil {
+		t.Fatalf("tools 不是合法 JSON: %v (%s)", err, tools)
+	}
+	if len(decoded) != 1 || decoded[0]["type"] != "function" {
+		t.Errorf("tools 形状被内核改写: %s", tools)
+	}
+	if tc := string(r.rawToolChoice()); tc != `"auto"` {
+		t.Errorf("tool_choice = %s, want \"auto\"", tc)
+	}
+	// 空值不应变成 "null" 出现在 body 里
+	var empty chatRequest
+	if got := empty.rawTools(); got != nil {
+		t.Errorf("无 tools 时应返回 nil, got %s", got)
+	}
+	if got := empty.rawToolChoice(); got != nil {
+		t.Errorf("无 tool_choice 时应返回 nil, got %s", got)
+	}
+}
+
+func boolPtr(v bool) *bool { return &v }

internal/gateway/ui_image_page_test.go

@@ -0,0 +1,195 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/provider"
+)
+
+// TestUI_ChatPageGroupsImageModels is a UI-contract guard for the bug that
+// started this: the chat page listed image models beside chat ones, and picking
+// one posted to /api/chat, which cannot serve it.
+//
+// The contract is asserted on the HTML source rather than by running a browser,
+// because the failure this must catch is a source-level one — an <option>
+// without `disabled`. A DOM assertion in a headless browser would pass right
+// up until someone removed the attribute, and then only if a CI machine had the
+// whole stack up.
+func TestUI_ChatPageGroupsImageModels(t *testing.T) {
+	ui := uiSource(t)
+
+	// The dropdown must read model_kinds, not the flat models list. Reading the
+	// flat list is what made chat and image models indistinguishable.
+	if !strings.Contains(ui, "model_kinds") {
+		t.Error("chat page never reads model_kinds; image models will be " +
+			"indistinguishable from chat models again")
+	}
+
+	// Image options must be rendered disabled. A plain String.replace over the
+	// joined HTML only disables the FIRST match — that bug shipped once here
+	// (one source disabled, five selectable) — so the assertion looks for the
+	// per-item map, not for a global replace.
+	if strings.Contains(ui, `imgOpts.join("").replace`) {
+		t.Error("chat page disables image options with String.replace over the " +
+			"joined HTML, which only rewrites the first option")
+	}
+	if !strings.Contains(ui, "disabled") {
+		t.Error("chat page never emits a disabled attribute for image models")
+	}
+}
+
+// TestUI_HasDedicatedImagePage pins the second half of the fix: image models
+// need somewhere they CAN be tested.
+func TestUI_HasDedicatedImagePage(t *testing.T) {
+	ui := uiSource(t)
+
+	for _, want := range []string{
+		`data-tab="image"`,       // sidebar entry
+		`id="tab-image"`,         // pane
+		`function renderImage`,   // renderer registered with refresh()
+		`tab === "image"`,        // dispatched
+		`/v1/images/generations`, // the endpoint the page actually calls
+	} {
+		if !strings.Contains(ui, want) {
+			t.Errorf("image page is missing %q", want)
+		}
+	}
+
+	// The page must be reachable from TABS; goTab iterates TABS to toggle pane
+	// visibility, so a pane outside that list is present but permanently hidden —
+	// which reads as "the page is blank" because the content is in the DOM.
+	if !strings.Contains(ui, `"image"`) {
+		t.Error(`"image" does not appear in the TABS list`)
+	}
+}
+
+// TestUI_ImagePageCallsTheImageEndpoint is the discriminator: a page that
+// renders an image result but posts to the chat endpoint would produce exactly
+// the original failure, just prettier.
+func TestUI_ImagePageCallsTheImageEndpoint(t *testing.T) {
+	ui := uiSource(t)
+
+	i := strings.Index(ui, "async function sendImage(")
+	if i < 0 {
+		t.Fatal("sendImage() not found")
+	}
+	// Bound the search to the function body rather than the whole file, so a
+	// /v1/images/generations mention elsewhere cannot satisfy this.
+	body := ui[i:]
+	if end := strings.Index(body, "\n      function "); end > 0 {
+		body = body[:end]
+	}
+	if end := strings.Index(body, "\n      async function "); end > 0 {
+		if s := body[:end]; strings.Contains(s, "/v1/images/generations") {
+			body = s
+		} else {
+			body = body[:end]
+		}
+	}
+	if !strings.Contains(body, "/v1/images/generations") {
+		t.Error("sendImage() does not post to /v1/images/generations")
+	}
+}
+
+// TestSourceStatus_ExposesModelKindsToTheUI is the server half of the same
+// contract: the field the UI reads must actually be serialized.
+func TestSourceStatus_ExposesModelKindsToTheUI(t *testing.T) {
+	st := provider.SourceStatus{
+		Name:   "vela",
+		Models: []string{"claude-sonnet-5.5", "gpt-image-2"},
+		ModelKinds: []provider.ModelKindInfo{
+			{ID: "claude-sonnet-5.5", Kind: "chat"},
+			{ID: "gpt-image-2", Kind: "image"},
+		},
+	}
+	b, err := json.Marshal(st)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got map[string]interface{}
+	if err := json.Unmarshal(b, &got); err != nil {
+		t.Fatal(err)
+	}
+	mk, ok := got["model_kinds"].([]interface{})
+	if !ok || len(mk) != 2 {
+		t.Fatalf("model_kinds = %v, want 2 entries; the UI cannot split the list without it", got["model_kinds"])
+	}
+	last := mk[1].(map[string]interface{})
+	if last["kind"] != "image" {
+		t.Errorf("model_kinds[1].kind = %v, want \"image\"", last["kind"])
+	}
+}
+
+// TestUI_BaseURLFieldExplainsWhereToCut keeps the second bug fixed: the URL
+// hint has to mention the concrete failure, not just say "enter the URL".
+func TestUI_BaseURLFieldExplainsWhereToCut(t *testing.T) {
+	ui := uiSource(t)
+	if !strings.Contains(ui, "urlHint") {
+		t.Fatal("no urlHint; the base_url field gives no guidance about /v1")
+	}
+	hint := ui[strings.Index(ui, "urlHint:"):]
+	if len(hint) > 700 {
+		hint = hint[:700]
+	}
+	// The trap is specific and reproducible: base_url ending in /v1 makes chat
+	// work and image generation 404. Saying only "enter the URL" reproduces the
+	// confusion the hint exists to remove.
+	for _, want := range []string{"/v1", "404"} {
+		if !strings.Contains(hint, want) {
+			t.Errorf("urlHint does not mention %q — the hint must name the concrete failure", want)
+		}
+	}
+}
+
+// TestImageEndpointComposesUnderRealConditions proves the hint describes a
+// reachable path rather than a plausible-sounding one.
+//
+// The case that matters: base_url already ends in /v1 (which is what an
+// operator pastes from an SDK example), so the RELATIVE image endpoint composes
+// to …/v1/images/generations and works, while the gateway's default ABSOLUTE
+// path would produce …/v1/v1/… and 404.
+//
+// The doubled form is served as a 404 here so the test would fail if the code
+// ever reverted to relying on the default.
+func TestImageEndpointComposesUnderRealConditions(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v1/images/generations":
+			_, _ = w.Write([]byte(`{"data":[{"url":"http://example.invalid/x.png"}]}`))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer up.Close()
+
+	// base_url as an operator types it: host root + /v1.
+	base := up.URL + "/v1"
+
+	// What discovery writes when base carries a version segment.
+	relative := base + "/images/generations"
+	resp, err := http.Post(relative, "application/json", strings.NewReader(`{"model":"m"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("POST %s = %d, want 200 — the form discovery writes must work", relative, resp.StatusCode)
+	}
+
+	// What the gateway's built-in default would produce for the same base_url.
+	doubled := base + "/v1/images/generations"
+	resp2, err := http.Post(doubled, "application/json", strings.NewReader(`{"model":"m"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp2.Body.Close()
+	if resp2.StatusCode != http.StatusNotFound {
+		t.Errorf("POST %s = %d, want 404 — this test's premise is that the "+
+			"default image path doubles the version segment; if the upstream ever "+
+			"accepts it, composeImageEndpoint's rule needs rechecking", doubled, resp2.StatusCode)
+	}
+}

internal/gateway/ui_image_wire_test.go

@@ -0,0 +1,109 @@
+package gateway
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestUI_ImageRequestSendsJSONBody pins the bug that made the image page
+// unusable in production while every DOM assertion still passed.
+//
+// The page called api(path, {method, body: {...}}). api() is
+// `(p, o) => fetch(p, o)` — o is the NATIVE fetch options object, so an object
+// in `body` is coerced by fetch into the literal text "[object Object]". The
+// gateway's json.Unmarshal rejected it, the response was a bare 400 with
+// lat_ms 0, and the page showed an error toast with no image and no usable
+// message.
+//
+// What made this expensive: the pane rendered, the model list had all six
+// models, the button enabled itself, the console was clean. Only the WIRE
+// format was wrong, and nothing in the DOM shows a wire format.
+func TestUI_ImageRequestSendsJSONBody(t *testing.T) {
+	body := sendImageSource(t)
+
+	if !strings.Contains(body, "JSON.stringify") {
+		t.Error("sendImage() does not JSON.stringify its body — fetch coerces " +
+			"the object to the text \"[object Object]\" and the gateway answers " +
+			"400 with lat_ms 0")
+	}
+	if !strings.Contains(body, `"Content-Type": "application/json"`) {
+		t.Error("sendImage() does not set Content-Type: application/json; the " +
+			"server sees text/plain and may refuse to parse the body")
+	}
+	// The literal defect, asserted so a later refactor reintroducing it fails
+	// here rather than in a browser.
+	if strings.Contains(body, "body: size ? {") || strings.Contains(body, "body: { model") {
+		t.Error("sendImage() still passes an object literal as the fetch body")
+	}
+}
+
+// TestUI_PostCallsSendJSONString guards every POST in the file, not just the
+// image one. The image page happened to be the only place that used api() with
+// an object body, so a check scoped to it would pass while another page had the
+// same defect.
+func TestUI_PostCallsSendJSONString(t *testing.T) {
+	ui := uiSource(t)
+	lines := strings.Split(ui, "\n")
+
+	var bad []string
+	for i, l := range lines {
+		if !strings.Contains(l, `method: "POST"`) {
+			continue
+		}
+		end := i + 8
+		if end > len(lines) {
+			end = len(lines)
+		}
+		window := strings.Join(lines[i:end], "\n")
+		if !strings.Contains(window, "body:") {
+			continue // no body at all: legitimate (a bare trigger)
+		}
+		if strings.Contains(window, "JSON.stringify") || strings.Contains(window, "FormData") {
+			continue
+		}
+		bad = append(bad, strings.TrimSpace(l))
+	}
+	if len(bad) > 0 {
+		t.Errorf("%d POST call(s) send a non-string body:", len(bad))
+		for _, b := range bad {
+			if len(b) > 100 {
+				b = b[:100]
+			}
+			t.Errorf("    %s", b)
+		}
+	}
+}
+
+// TestUI_ImagePageReportsServerErrors pins that a failed generation surfaces
+// something actionable. The production bug produced a bare "400 Bad Request"
+// toast; the requirement is that the gateway's own message reaches the page.
+func TestUI_ImagePageReportsServerErrors(t *testing.T) {
+	body := sendImageSource(t)
+
+	// api() throws Error(j.error.message) on !r.ok, so appending e.message is
+	// what puts the explanation on screen.
+	if !strings.Contains(body, "e && e.message") {
+		t.Error("sendImage() does not surface the server's error message; a " +
+			"failure would show a bare status with no cause")
+	}
+	// And a success carrying no image must not look like a success.
+	if !strings.Contains(body, "imgNoData") {
+		t.Error("sendImage() does not warn when the upstream returns no image")
+	}
+}
+
+// sendImageSource returns the text of sendImage() only, so an assertion cannot
+// be satisfied by a string that happens to appear elsewhere in the file.
+func sendImageSource(t *testing.T) string {
+	t.Helper()
+	ui := uiSource(t)
+	i := strings.Index(ui, "async function sendImage(")
+	if i < 0 {
+		t.Fatal("sendImage() not found")
+	}
+	body := ui[i:]
+	if end := strings.Index(body[1:], "\n      function "); end > 0 {
+		body = body[:end+1]
+	}
+	return body
+}

internal/gateway/ui_key_auto_test.go

@@ -0,0 +1,439 @@
+package gateway
+
+import (
+	"os"
+	"strings"
+	"testing"
+)
+
+// Per-key AUTO chain UI contract.
+//
+// The per-key editor deliberately REUSES the global lane canvas instead of
+// shipping a second, simpler editor. That is a design choice with teeth: a
+// reduced copy would lack drag ordering and per-slot quota/period/hours, and
+// the two would drift apart silently. These checks pin the sharing, because
+// the only symptom of un-sharing is "the per-key editor got worse" — no error,
+// just a smaller feature set that nobody notices until they use it.
+func TestUIPerKeyAutoWiring(t *testing.T) {
+	src := uiSource(t)
+
+	// The admin key list must offer the editor, otherwise the whole feature is
+	// unreachable from the UI. Assert the *call site* inside keyCanvasHtml,
+	// not just that the functions exist: removing the button leaves every
+	// definition intact, and that is exactly the regression to catch.
+	kcStart := strings.Index(src, "function keyCanvasHtml")
+	if kcStart < 0 {
+		t.Fatal("keyCanvasHtml is gone — cannot check the AUTO chain button")
+	}
+	kcEnd := strings.Index(src[kcStart:], "\n      }")
+	if kcEnd < 0 {
+		t.Fatal("could not delimit keyCanvasHtml")
+	}
+	if !strings.Contains(src[kcStart:kcStart+kcEnd], "openKeyAutoModal(") {
+		t.Fatal("per-key AUTO chain button is missing from the key card header")
+	}
+	for _, fn := range []string{
+		"function openKeyAutoModal",
+		"function keyAutoInherit",
+		"function keyAutoSave",
+		"function keyAutoClose",
+	} {
+		if !strings.Contains(src, fn) {
+			t.Fatalf("UI is missing %s — the editor cannot function", fn)
+		}
+	}
+}
+
+// The per-key editor must render the SAME canvas the global editor uses.
+// If this regresses, the per-key chain silently loses drag ordering and
+// per-slot quota/period/hours while still looking like it works.
+func TestUIPerKeyAutoReusesGlobalCanvas(t *testing.T) {
+	src := uiSource(t)
+
+	// One canvas implementation, two mount ids.
+	if !strings.Contains(src, "function renderSortEditor(") {
+		t.Fatal("no shared canvas renderer — the per-key editor must reuse renderSortEditor")
+	}
+	if !strings.Contains(src, `canvasId: "scr-canvas"`) {
+		t.Fatal("the global editor no longer mounts the shared canvas")
+	}
+	if !strings.Contains(src, `canvasId: "key-auto-canvas"`) {
+		t.Fatal("the per-key editor does not mount the shared canvas")
+	}
+	// Canvas lookup must consider both ids, or drag/drop silently no-ops on
+	// whichever surface is not #scr-canvas.
+	if !strings.Contains(src, "function sortCanvasEl(") ||
+		!strings.Contains(src, `getElementById("key-auto-canvas")`) {
+		t.Fatal("sortCanvasEl must resolve both canvas ids")
+	}
+	// The rules<->lanes conversion is shared too, so the per-key editor sends
+	// exactly the fields the server understands (quota/period/hours included).
+	for _, fn := range []string{"function buildLanes(", "function lanesToRules("} {
+		if !strings.Contains(src, fn) {
+			t.Fatalf("%s must be shared between the global and per-key editors", fn)
+		}
+	}
+}
+
+// Saving must route through persistAuto's scope switch, not a hand-rolled
+// request, or the two editors will serialize rules differently.
+func TestUIPerKeyAutoSavesThroughScope(t *testing.T) {
+	src := uiSource(t)
+	if !strings.Contains(src, "sortState.scope = key") {
+		t.Fatal("openKeyAutoModal must set sortState.scope so the save targets the key")
+	}
+	if !strings.Contains(src, "sortState.scope = null") {
+		t.Fatal("closing the modal must clear sortState.scope, or the GLOBAL editor would save to this key")
+	}
+	if !strings.Contains(src, "if (sortState.scope)") {
+		t.Fatal("persistAuto must branch on sortState.scope")
+	}
+	if !strings.Contains(src, `"/auto"`) {
+		t.Fatal("per-key editor must call the /api/keys/{key}/auto endpoint")
+	}
+	// An empty lane list is how "inherit the global chain" is expressed.
+	if !strings.Contains(src, "JSON.stringify({ auto: lanesToRules(sortState.lanes) })") {
+		t.Fatal("keyAutoSave must PUT lanesToRules(sortState.lanes) verbatim")
+	}
+}
+
+// Both languages need the strings, otherwise a non-Chinese admin sees raw
+// i18n keys in the modal.
+func TestUIPerKeyAutoHasBothLanguages(t *testing.T) {
+	src := uiSource(t)
+	for _, k := range []string{
+		"kAutoChainTitle", "kAutoChainHint", "kAutoChainInherit",
+		"kAutoChainEmpty", "kAutoChainGlobal", "kAutoChainBad",
+	} {
+		n := strings.Count(src, k+`: "`)
+		if n < 2 {
+			t.Fatalf("i18n key %s defined %d time(s), want both zh and en", k, n)
+		}
+	}
+}
+
+// The statistics dashboard must print the window it actually covers. Calendar
+// periods are not nested — "this week" starts on its Monday and therefore
+// reaches into the previous month early in a month, making its total LARGER
+// than "this month". That is correct arithmetic, but without the dates on
+// screen it reads as a bug, and it has been reported as one.
+func TestUIStatsShowsPeriodWindow(t *testing.T) {
+	src := uiSource(t)
+	if !strings.Contains(src, `id="period-range"`) {
+		t.Fatal("the stats dashboard has no element for the period window range")
+	}
+	if n := strings.Count(src, "periodRange"); n < 2 {
+		t.Fatalf("i18n key periodRange appears %d time(s), want both zh and en", n)
+	}
+}
+
+// Both canvases can exist in the DOM at once: the modal is mounted on top of
+// the settings tab, whose #scr-canvas stays alive behind it. Resolving the
+// canvas by document order (whichever comes first) painted the per-key chain
+// into the hidden background tab and left the dialog empty — and worse,
+// overwrote the global chain the user returns to. The lookup must follow
+// sortState.scope instead.
+func TestUISortCanvasResolvesByScope(t *testing.T) {
+	src := uiSource(t)
+	el := readFuncBody(t, src, "sortCanvasEl")
+	if !strings.Contains(el, "if (sortState.scope)") {
+		t.Fatal("sortCanvasEl must branch on sortState.scope, not document order")
+	}
+	if strings.Contains(el, `getElementById("scr-canvas")`) &&
+		strings.Index(el, `getElementById("scr-canvas")`) < strings.Index(el, "sortState.scope") {
+		t.Fatal("sortCanvasEl returns #scr-canvas before checking scope — the " +
+			"per-key editor paints into the global canvas")
+	}
+}
+
+// A successful save must close the dialog even when the slots do not resolve.
+// The write DID succeed; leaving the modal open makes the user click Save
+// again and again. This regressed once: the !slots branch returned early.
+func TestUIKeyAutoSaveClosesOnSuccess(t *testing.T) {
+	src := uiSource(t)
+	body := readFuncBody(t, src, "keyAutoSave")
+	// keyAutoClose() must be reached on every success path, not only in the
+	// happy branch.
+	iClose := strings.Index(body, "keyAutoClose()")
+	if iClose < 0 {
+		t.Fatal("keyAutoSave never calls keyAutoClose — the dialog would stay open")
+	}
+	iToast := strings.Index(body, "toast(")
+	if iToast >= 0 && iClose < iToast {
+		t.Fatal("keyAutoClose is called before the toast; the warning must be " +
+			"shown first or it is lost with the modal")
+	}
+	// An early `return` inside a branch bypasses keyAutoClose entirely, so
+	// merely finding the call is not enough: every `return` that sits between
+	// entering the try block and the close call is a path that leaves the
+	// dialog open after a successful write.
+	tryStart := strings.Index(body, "try {")
+	seg := body[tryStart:iClose]
+	for _, line := range strings.Split(seg, "\n") {
+		trimmed := strings.TrimSpace(line)
+		if trimmed == "return;" || strings.HasPrefix(trimmed, "return ") {
+			t.Fatal("a `return` before keyAutoClose leaves the dialog open after " +
+				"a successful save: " + trimmed)
+		}
+	}
+}
+
+// readFuncBody returns the source of a top-level function by name, using
+// brace matching so nested blocks do not truncate it.
+func readFuncBody(t *testing.T, src, name string) string {
+	t.Helper()
+	idx := strings.Index(src, "function "+name+"(")
+	if idx < 0 {
+		t.Fatalf("function %s not found", name)
+	}
+	start := strings.Index(src[idx:], "{")
+	if start < 0 {
+		t.Fatalf("function %s has no body", name)
+	}
+	i := idx + start
+	depth := 0
+	for j := i; j < len(src); j++ {
+		switch src[j] {
+		case '{':
+			depth++
+		case '}':
+			depth--
+			if depth == 0 {
+				return src[i : j+1]
+			}
+		}
+	}
+	t.Fatalf("function %s body is unterminated", name)
+	return ""
+}
+
+// The per-key dialog must not share the app-wide #modal-wrap id.
+//
+// Every dialog in this file uses that id, and more than one can be open at
+// once (the "add slot" picker opens ON TOP of this one). Handlers resolve
+// their own dialog via `.closest()` from their button, which works — but this
+// dialog is also looked up from the CANVAS side (sortCanvasEl, keyAutoClose,
+// keyAutoInherit), where there is no button to walk up from.
+// getElementById("#modal-wrap") then returns whichever dialog comes first in
+// the document, i.e. not the one being edited: Cancel closed the wrong box and
+// the canvas could resolve to the picker's.
+func TestUIKeyAutoModalHasOwnID(t *testing.T) {
+	src := uiSource(t)
+	body := readFuncBody(t, src, "openKeyAutoModal")
+	if !strings.Contains(body, `wrap.id = "key-auto-modal"`) {
+		t.Fatal("openKeyAutoModal must give its dialog its own id, not the shared " +
+			"#modal-wrap (another dialog may be open on top of it)")
+	}
+	close := readFuncBody(t, src, "keyAutoClose")
+	if strings.Contains(close, `"modal-wrap"`) {
+		t.Fatal("keyAutoClose resolves the shared #modal-wrap; it would close " +
+			"whichever dialog comes first in the document")
+	}
+	if !strings.Contains(close, `"key-auto-modal"`) {
+		t.Fatal("keyAutoClose must resolve the per-key dialog by its own id")
+	}
+}
+
+// Editing a chain inside the per-key dialog must NOT persist it. That dialog
+// has Save and Cancel, so auto-writing on add/drag/delete meant the change was
+// already committed before Save and Cancel could not undo it.
+//
+// The global page is the opposite: it has always applied edits immediately.
+func TestUIKeyAutoEditsDoNotAutoPersist(t *testing.T) {
+	src := uiSource(t)
+	body := readFuncBody(t, src, "afterChainEdit")
+	if !strings.Contains(body, "if (sortState.scope)") {
+		t.Fatal("afterChainEdit must branch on sortState.scope")
+	}
+	// The scoped branch must return before reaching persistAuto.
+	iScope := strings.Index(body, "if (sortState.scope)")
+	iReturn := strings.Index(body[iScope:], "return")
+	iPersist := strings.Index(body, "persistAuto()")
+	if iScope < 0 || iReturn < 0 || iPersist < 0 {
+		t.Fatal("afterChainEdit no longer has the expected shape")
+	}
+	if iScope+iReturn > iPersist {
+		t.Fatal("the scoped branch reaches persistAuto() — a per-key edit would " +
+			"be written before Save")
+	}
+	// And it must not call itself (a regex-driven rewrite did exactly that).
+	if strings.Contains(body, "afterChainEdit()") {
+		t.Fatal("afterChainEdit calls itself: infinite recursion")
+	}
+	// Every in-canvas edit goes through it, so none of them bypass the check.
+	for _, fn := range []string{"scrAddFromForm", "sortScopeSave", "scrDelSlot"} {
+		b := readFuncBody(t, src, fn)
+		if strings.Contains(b, "persistAuto()") {
+			t.Fatalf("%s calls persistAuto() directly, bypassing afterChainEdit", fn)
+		}
+	}
+}
+
+// The stats range line formats its end date from the LAST bucket label.
+// Bucket labels are hourly for the day view ("2026-10-03T09") and daily for
+// week/month ("2026-10-01"), so blindly appending "T00:00:00Z" to the hourly
+// form builds "2026-10-03T09T00:00:00Z" — Date rejects it, toISOString()
+// throws, and the assignment is lost. The symptom is an empty range line on
+// exactly the view people look at most, with no error anywhere.
+func TestUIStatsPeriodRangeHandlesHourlyBuckets(t *testing.T) {
+	src := uiSource(t)
+	el := readFuncBody(t, src, "paintStats")
+	// It must branch on the label shape instead of concatenating blindly.
+	if !strings.Contains(el, `T\d{2}$`) && !strings.Contains(el, `\d{2}-\d{2}T\d{2}`) {
+		t.Fatal("paintStats must recognise the hourly bucket label shape " +
+			"(YYYY-MM-DDTHH); concatenating a time onto it yields an invalid Date")
+	}
+	// And it must not contain the naive form that broke it.
+	if strings.Contains(el, `bucket + "T00:00:00Z"`) {
+		t.Fatal("paintStats still appends T00:00:00Z to the bucket label — " +
+			"this throws on the day view and leaves the range line empty")
+	}
+}
+
+// A key with no chain of its own must open on a COPY OF THE GLOBAL CHAIN, not
+// on an empty canvas. The realistic admin task is "this user gets the global chain
+// minus the sources they must not hit"; starting blank forces retyping every
+// tier against a 226-model picker. This is deliberately NOT the discovery
+// fallback (which dumps every model of every source) — it is the same rules
+// the global chain actually runs.
+func TestUIKeyAutoSeedsFromGlobalChain(t *testing.T) {
+	src := uiSource(t)
+	body := readFuncBody(t, src, "openKeyAutoModal")
+	if !strings.Contains(body, `api("/api/auto")`) {
+		t.Fatal("openKeyAutoModal must read the global chain to seed the editor")
+	}
+	if !strings.Contains(body, "seededFromGlobal") {
+		t.Fatal("the editor must remember whether its content came from the " +
+			"global chain, so the notice can say so")
+	}
+	// And it must not fall back to the discovery path that fills 226 models.
+	if strings.Contains(body, `buildLanes(idx, [], `) {
+		t.Fatal("seeding from an empty rule list triggers buildLanes discovery")
+	}
+}
+
+// The notice must distinguish "already this key's chain" from "a copy that
+// only becomes its chain on Save" — those two look identical on the canvas and
+// behave very differently if the admin hits Cancel.
+func TestUIKeyAutoNoticeDistinguishesSavedFromCopied(t *testing.T) {
+	src := uiSource(t)
+	body := readFuncBody(t, src, "keyAutoInheritNotice")
+	for _, k := range []string{"kAutoChainSeeded", "kAutoChainNewOwn", "kAutoChainWillInherit"} {
+		if !strings.Contains(body, k) {
+			t.Fatalf("the notice never uses %s — a copied chain and a saved one "+
+				"would look the same", k)
+		}
+	}
+	// Both languages, or a non-Chinese admin sees raw keys.
+	for _, k := range []string{"kAutoChainSeeded", "kAutoChainNewOwn", "kAutoChainWillInherit"} {
+		if n := strings.Count(src, k+`: "`); n < 2 {
+			t.Fatalf("i18n key %s defined %d time(s), want both zh and en", k, n)
+		}
+	}
+}
+
+// A key whose model scope excludes AUTO cannot use AUTO at all — the scope
+// filter runs before the chain — yet the editor happily lets an admin
+// configure a per-key chain. That combination is easy to create by accident
+// and produces a 403 with no visible cause, so the dialog must say so.
+//
+// We deliberately do NOT widen the scope automatically: silently granting a
+// model the operator did not ask for is worse than a loud warning.
+func TestUIKeyAutoWarnsOnScopeConflict(t *testing.T) {
+	src := uiSource(t)
+	if !strings.Contains(src, "function keyAutoScopeWarning(") {
+		t.Fatal("no scope-conflict check in the per-key editor")
+	}
+	body := readFuncBody(t, src, "keyAutoScopeWarning")
+	if !strings.Contains(body, "AUTO") {
+		t.Fatal("keyAutoScopeWarning must test whether the scope lists AUTO")
+	}
+	// It must not write anything back to the key: reporting only.
+	for _, forbidden := range []string{"PUT", "POST", "fetch(", "api("} {
+		if strings.Contains(body, forbidden) {
+			t.Fatalf("keyAutoScopeWarning calls %s — it must only report the "+
+				"conflict, never widen the scope itself", forbidden)
+		}
+	}
+	// Empty scope means unrestricted, so no warning is correct there.
+	if !strings.Contains(body, "if (!models.length)") {
+		t.Fatal("an unrestricted key (no models) must not be warned")
+	}
+	// Both languages.
+	if n := strings.Count(src, "kAutoChainScopeWarn"); n < 2 {
+		t.Fatalf("i18n key kAutoChainScopeWarn appears %d time(s), want zh+en", n)
+	}
+}
+
+// The scope conflict for AUTO must be reported by checkQuota, because that is
+// the check an AUTO request actually reaches first (it runs before the chain is
+// consulted). Putting the message on the later path made it dead code — the
+// request was rejected one check earlier with the generic "not allowed for this
+// key" wording, which sends an operator looking for a quota that isn't the
+// problem.
+func TestAutoScopeConflictMessageComesFromCheckQuota(t *testing.T) {
+	b, err := os.ReadFile("chat.go")
+	if err != nil {
+		t.Fatalf("read chat.go: %v", err)
+	}
+	src := string(b)
+	i := strings.Index(src, "func (g *Gateway) checkQuota(")
+	if i < 0 {
+		t.Fatal("checkQuota not found")
+	}
+	// Body = up to the next top-level func.
+	rest := src[i:]
+	if j := strings.Index(rest[1:], "\nfunc "); j > 0 {
+		rest = rest[:j+1]
+	}
+	if !strings.Contains(rest, "isAuto(model)") {
+		t.Fatal("checkQuota must special-case AUTO: a scope that omits AUTO is " +
+			"rejected there before anything else looks at the chain")
+	}
+	if !strings.Contains(rest, "not in this key's model scope") {
+		t.Fatal("checkQuota's AUTO branch must name the scope as the cause")
+	}
+}
+
+// renderSortEditor replaces its mount's innerHTML wholesale. The global
+// priority page used to hand it the .card that also holds the h2 toolbar, so
+// every open deleted the save/reset/add buttons: the lanes still rendered and
+// the drag worked, but there was no way to persist a change. The per-key modal
+// never showed this because its mount held only the canvas.
+//
+// The regression is invisible in the DOM when you only count lane elements, so
+// this asserts on the structure: the toolbar lives in a sibling of the mount,
+// not inside it.
+func TestGlobalSortToolbarSurvivesCanvasRender(t *testing.T) {
+	src := uiSource(t)
+
+	// The global editor must mount into a dedicated host, not the .card.
+	i := strings.Index(src, `$("#tab-sort").innerHTML = `)
+	if i < 0 {
+		t.Fatal("global sort markup not found")
+	}
+	chunk := src[i:min(len(src), i+1400)]
+	if !strings.Contains(chunk, `id="sort-mount"`) {
+		t.Fatal("global sort page must provide a dedicated #sort-mount for the canvas")
+	}
+	// The mount handed to renderSortEditor must be that host.
+	if !strings.Contains(src, `renderSortEditor($("#sort-mount")`) {
+		t.Fatal("renderSortEditor must be given #sort-mount; passing the .card " +
+			"wipes the h2 toolbar that holds the save button")
+	}
+	// The toolbar (save in particular) must be OUTSIDE #sort-mount.
+	close := strings.Index(chunk, "</div>`;")
+	if close < 0 {
+		t.Fatal("malformed #tab-sort template")
+	}
+	head := chunk[:close]
+	mountAt := strings.Index(head, `id="sort-mount"`)
+	saveAt := strings.Index(head, `onclick="saveSort()"`)
+	if saveAt < 0 {
+		t.Fatal("the global sort page has no save button at all")
+	}
+	if mountAt >= 0 && saveAt > mountAt {
+		t.Fatal("the save button sits inside the canvas mount and will be " +
+			"erased the moment the canvas renders")
+	}
+}

internal/gateway/ui_plugin_test.go

@@ -388,3 +388,26 @@ func TestPluginSidebarEntryCarriesATitle(t *testing.T) {
 	t.Errorf("plugin sidebar button renders the icon only; a label is required. "+
 		"got: %.200s", stmt)
 }
+
+// TestPluginPageHasAccessEditor: the plugin page must let an operator configure
+// endpoint access. The APIs existed with zero UI entry point, which meant the
+// only way to change who could reach a plugin endpoint was to hand-edit
+// config.yaml — the thing the UI is meant to replace.
+func TestPluginPageHasAccessEditor(t *testing.T) {
+	src := uiSource(t)
+	for _, needle := range []string{
+		`renderPluginAccess`, // called after rendering the plugin list
+		`savePluginAccess`,   // wired to a button
+		`/access`,            // hits the per-endpoint access API
+		`id="pl-access"`,     // container exists
+		`http_access`,        // reads the table from /api/plugins
+		`data-ep=`,           // one input per endpoint
+		`data-plvis=`,        // one visibility selector per plugin
+		`/http/*/access`,     // whole-plugin visibility shares the access API
+		`plVisCell`,          // renders the declared-vs-override distinction
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("plugin page lacks %q", needle)
+		}
+	}
+}

internal/lua/adapter_qwen38_test.go

@@ -0,0 +1,323 @@
+package lua
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// The fixtures below are REAL responses captured from
+// https://discovery-api.intern-ai.org.cn/v1 with qwen3.8-27b — not hand-written
+// approximations. Two of them exist only because the endpoint behaves in a way a
+// stock OpenAI adapter gets wrong, and neither is guessable from the schema.
+
+// reasoningDominatesTheBudget is the fact that made a dedicated adapter
+// necessary: a one-sentence visual answer came back as completion=303 with
+// reasoning=295. 97% of the budget went to thinking and only 8 tokens reached
+// content. Handing the upstream's content straight through yields what looks
+// like an empty answer from a perfectly healthy request.
+const reasoningDominatesTheBudget = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":"\n\n这张图是白色的。",
+    "reasoning_content":"分析请求…分析图片…颜色是纯白"},"logprobs":null,
+    "finish_reason":"stop"}],
+  "created":1791179068,"model":"qwen3.8-27b-fp8","object":"chat.completion",
+  "usage":{"prompt_tokens":125,"total_tokens":428,"completion_tokens":303,
+    "prompt_tokens_details":null,
+    "completion_tokens_details":{"reasoning_tokens":295}}
+}`
+
+// truncatedReasoningOnly is what a small max_tokens produces: HTTP 200, content
+// null, finish_reason "length". It is NOT an upstream failure.
+const truncatedReasoningOnly = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":null,
+    "reasoning_content":"思考过程：1. **分析请求**…图片的颜色是纯","refusal":null},
+    "finish_reason":"length"}],
+  "model":"qwen3.8-27b-fp8","object":"chat.completion",
+  "usage":{"prompt_tokens":127,"total_tokens":190,"completion_tokens":63}
+}`
+
+const qwen38ToolCallResponse = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":"",
+    "reasoning_content":"用户询问北京天气。我需要调用 get_weather 函数，参数为城市 \"北京\"。",
+    "tool_calls":[{"id":"chatcmpl-tool-afc0299b5810e044","type":"function",
+      "function":{"name":"get_weather","arguments":"{\"city\": \"北京\"}"}}]},
+    "finish_reason":"tool_calls"}],
+  "model":"qwen3.8-27b-fp8",
+  "usage":{"prompt_tokens":305,"total_tokens":353,"completion_tokens":48,
+    "completion_tokens_details":{"reasoning_tokens":20}}
+}`
+
+func qwen38VM(t *testing.T) *VM {
+	t.Helper()
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("adapter VM: %v", err)
+	}
+	t.Cleanup(vm.Stop)
+	return vm
+}
+
+func qwen38Response(t *testing.T, vm *VM, raw string) map[string]interface{} {
+	t.Helper()
+	out, err := vm.Transform("qwen38", "transform_response", raw)
+	if err != nil {
+		t.Fatalf("transform_response: %v", err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("adapter returned non-JSON: %v\n%s", err, out)
+	}
+	return m
+}
+
+func qwen38Stream(t *testing.T, vm *VM, raw string) map[string]interface{} {
+	t.Helper()
+	out, err := vm.Transform("qwen38", "transform_stream_chunk", raw)
+	if err != nil {
+		t.Fatalf("transform_stream_chunk: %v", err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("adapter returned non-JSON: %v\n%s", err, out)
+	}
+	return m
+}
+
+// TestQwen38KeepsReasoningSeparate is the adapter's reason for existing.
+// Folding reasoning into content would render the model's private deliberation as
+// its answer; dropping it makes a reasoning model look like it returned nothing.
+func TestQwen38KeepsReasoningSeparate(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, reasoningDominatesTheBudget)
+
+	content, _ := u["content"].(string)
+	reasoning, _ := u["reasoning_content"].(string)
+
+	if reasoning == "" {
+		t.Fatal("reasoning_content was dropped — a reasoning model then looks like it returned nothing")
+	}
+	if strings.Contains(content, "分析请求") {
+		t.Errorf("reasoning was folded into content: %q", content)
+	}
+	if !strings.Contains(content, "白色的") {
+		t.Errorf("content = %q, want the answer text", content)
+	}
+	if reasoning == content {
+		t.Error("content and reasoning_content are identical; they are disjoint upstream")
+	}
+}
+
+// TestQwen38ReportsTruncationRatherThanFailing: content=null with
+// finish_reason="length" is a truncated reasoning response, not a broken
+// upstream. Losing the finish_reason makes it indistinguishable from a failure,
+// and losing the partial reasoning throws away the only evidence of what happened.
+func TestQwen38ReportsTruncationRatherThanFailing(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, truncatedReasoningOnly)
+
+	if u["finish_reason"] != "length" {
+		t.Errorf("finish_reason = %v, want length", u["finish_reason"])
+	}
+	if reasoning, _ := u["reasoning_content"].(string); reasoning == "" {
+		t.Error("the partial reasoning was discarded")
+	}
+	if content, _ := u["content"].(string); content != "" {
+		t.Errorf("content = %q, want empty string for a null upstream content", content)
+	}
+}
+
+// TestQwen38CarriesReasoningTokensWithoutInflatingTotals: reasoning_tokens is a
+// SUBSET of completion_tokens. Adding them double-counts every reasoning request,
+// and billing prices the pair separately.
+func TestQwen38CarriesReasoningTokensWithoutInflatingTotals(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, reasoningDominatesTheBudget)
+
+	tu, ok := u["token_usage"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("token_usage = %#v", u["token_usage"])
+	}
+	completion, _ := tu["completion"].(float64)
+	total, _ := tu["total"].(float64)
+
+	// The NESTING is the point, not the presence of the field.
+	//
+	// internal/gateway/chat.go reads reasoning in exactly one place:
+	//   resp.TokenUsage.CompletionTokensDetails.ReasoningTokens
+	// An adapter that emits reasoning_tokens at the TOP of token_usage type-checks
+	// fine and is then dropped at the one spot that could have kept it — measured
+	// on the live gateway: the model reported completion=88 and the gateway's
+	// response carried reasoning=null.
+	//
+	// So this asserts the path, not merely that a number appears somewhere.
+	det, ok := tu["completion_tokens_details"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("token_usage has no completion_tokens_details: %#v — chat.go reads "+
+			"reasoning from there and nowhere else, so the number is discarded", tu)
+	}
+	reasoning, has := det["reasoning_tokens"].(float64)
+	if !has {
+		t.Fatal("completion_tokens_details.reasoning_tokens missing — billing cannot price reasoning separately")
+	}
+	// Relational, not a presence check. The first version of this test only
+	// asserted the field existed, and a mutation setting it to 0 passed: 0 is
+	// present. The upstream value is known (295), so the exact number is
+	// asserted — a zero silently misprices every reasoning request.
+	if reasoning != 295 {
+		t.Errorf("reasoning_tokens = %v, want 295 (the upstream value)", reasoning)
+	}
+	if reasoning >= completion {
+		t.Errorf("reasoning_tokens (%v) must be a SUBSET of completion_tokens (%v)", reasoning, completion)
+	}
+	if completion != 303 || total != 428 {
+		t.Errorf("usage was altered: completion=%v total=%v, want 303/428", completion, total)
+	}
+}
+
+// TestQwen38DecodesToolCallArguments: the upstream sends arguments as a JSON
+// STRING, like every other OpenAI-compatible endpoint. Passing it through
+// undecoded leaves the caller to parse it.
+func TestQwen38DecodesToolCallArguments(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, qwen38ToolCallResponse)
+
+	tcs, ok := u["tool_calls"].([]interface{})
+	if !ok || len(tcs) != 1 {
+		t.Fatalf("tool_calls = %#v", u["tool_calls"])
+	}
+	tc, _ := tcs[0].(map[string]interface{})
+	if tc["name"] != "get_weather" {
+		t.Errorf("name = %v", tc["name"])
+	}
+	args, ok := tc["arguments"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("arguments = %#v — the upstream sends a JSON string and it must be decoded", tc["arguments"])
+	}
+	if args["city"] != "北京" {
+		t.Errorf("arguments.city = %v", args["city"])
+	}
+}
+
+// TestQwen38StreamKeepsReasoningSeparate: measured 20 reasoning chunks against 1
+// content chunk on a one-sentence answer. Folding them would make the client
+// render the deliberation as the reply.
+func TestQwen38StreamKeepsReasoningSeparate(t *testing.T) {
+	vm := qwen38VM(t)
+	const reasoningChunk = `{"choices":[{"index":0,"delta":{"reasoning_content":"思考过程：1. **分析请求**"},
+	  "finish_reason":null}],"model":"qwen3.8-27b-fp8"}`
+	const contentChunk = `{"choices":[{"index":0,"delta":{"content":"递归就是函数调用自身"},
+	  "finish_reason":null}],"model":"qwen3.8-27b-fp8"}`
+	const usageChunk = `{"choices":[],"usage":{"prompt_tokens":64,"total_tokens":132,
+	  "completion_tokens":68,"completion_tokens_details":{"reasoning_tokens":62}}}`
+
+	a := qwen38Stream(t, vm, reasoningChunk)
+	b := qwen38Stream(t, vm, contentChunk)
+	c := qwen38Stream(t, vm, usageChunk)
+
+	if content, mixed := a["content"].(string); mixed && content != "" {
+		t.Errorf("a reasoning chunk leaked into content: %q", content)
+	}
+	if reasoning, _ := a["reasoning_content"].(string); !strings.Contains(reasoning, "思考过程") {
+		t.Errorf("reasoning chunk lost its text: %#v", a)
+	}
+	if content, _ := b["content"].(string); !strings.Contains(content, "递归") {
+		t.Errorf("content chunk lost its text: %#v", b)
+	}
+	if c["usage"] == nil {
+		t.Errorf("the terminal usage chunk was dropped: %#v", c)
+		return
+	}
+	u, _ := c["usage"].(map[string]interface{})
+	det, _ := u["completion_tokens_details"].(map[string]interface{})
+	if det == nil || det["reasoning_tokens"] == nil {
+		t.Errorf("usage chunk did not nest reasoning under completion_tokens_details: %#v — "+
+			"chat.go reads it there, so a top-level field is discarded", u)
+	}
+}
+
+// TestQwen38ReportsMediaFailureCode: a remote image URL times out upstream and
+// comes back as media_fetch_failed. Surfacing only the message makes an image
+// fetch problem look like a model problem.
+func TestQwen38ReportsMediaFailureCode(t *testing.T) {
+	vm := qwen38VM(t)
+	body := `{"error":{"message":"image fetch timed out","type":"media_fetch_failed",
+	  "param":null,"code":"media_fetch_failed","trace_id":"f3ca4aa"}}`
+	reason, ok, err := vm.TransformError("qwen38", 400, body)
+	if err != nil {
+		t.Fatalf("TransformError: %v", err)
+	}
+	if !ok {
+		t.Fatal("the media failure was swallowed")
+	}
+	if !strings.Contains(reason, "image fetch timed out") || !strings.Contains(reason, "media_fetch_failed") {
+		t.Errorf("reason = %q; it must carry both the text and the code", reason)
+	}
+}
+
+// TestQwen38InvalidKeyError: the endpoint answers a bad key with 401 and
+// code "invalid_api_key". That one must NOT be annotated with its own code — it
+// would render as "invalid API key (invalid_api_key)".
+func TestQwen38InvalidKeyError(t *testing.T) {
+	vm := qwen38VM(t)
+	body := `{"error":{"message":"invalid API key","type":"invalid_request_error",
+	  "param":null,"code":"invalid_api_key"}}`
+	reason, ok, err := vm.TransformError("qwen38", 401, body)
+	if err != nil || !ok {
+		t.Fatalf("TransformError: ok=%v err=%v", ok, err)
+	}
+	if reason != "invalid API key" {
+		t.Errorf("reason = %q, want the bare message", reason)
+	}
+}
+
+// TestQwen38SuppliesReasoningOnToolCallTurns: the upstream requires an assistant
+// message carrying tool_calls to echo reasoning_content. OpenAI-shaped clients
+// never send it, so without this the SECOND turn of any tool-using conversation
+// is rejected.
+func TestQwen38SuppliesReasoningOnToolCallTurns(t *testing.T) {
+	vm := qwen38VM(t)
+	const req = `{"model":"qwen3.8-27b","messages":[
+	  {"role":"user","content":"北京天气"},
+	  {"role":"assistant","content":"","tool_calls":[{"id":"c1","type":"function",
+	    "function":{"name":"get_weather","arguments":"{\"city\":\"北京\"}"}}]},
+	  {"role":"tool","tool_call_id":"c1","content":"晴 25C"}]}`
+	out, err := vm.Transform("qwen38", "transform_request", req)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+	var decoded map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &decoded); err != nil {
+		t.Fatalf("request transform did not return JSON: %v\n%s", err, out)
+	}
+	msgs, _ := decoded["messages"].([]interface{})
+	if len(msgs) < 2 {
+		t.Fatalf("messages = %#v", decoded["messages"])
+	}
+	assistant, _ := msgs[1].(map[string]interface{})
+	if _, ok := assistant["reasoning_content"]; !ok {
+		t.Errorf("assistant turn with tool_calls has no reasoning_content; the upstream "+
+			"rejects the next turn. Got %#v", assistant)
+	}
+}
+
+// TestQwen38PassesImagesThrough: this adapter's job is reasoning handling, not
+// media translation. A multimodal content array must survive untouched — an
+// adapter that re-encodes it is the kind of change that silently breaks images
+// while every text test still passes.
+func TestQwen38PassesImagesThrough(t *testing.T) {
+	vm := qwen38VM(t)
+	const req = `{"model":"qwen3.8-27b","messages":[{"role":"user","content":[
+	  {"type":"text","text":"这张图是什么颜色？"},
+	  {"type":"image_url","image_url":{"url":"data:image/png;base64,iVBORw0KGgo="}}]}],
+	  "max_tokens":600}`
+	out, err := vm.Transform("qwen38", "transform_request", req)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+	if !strings.Contains(out, "image_url") || !strings.Contains(out, "data:image/png;base64,") {
+		t.Errorf("the image part was lost or rewritten:\n%s", out)
+	}
+	if strings.Contains(out, "base64,iVBORw0KGgo=") == false {
+		t.Errorf("the base64 payload was altered:\n%s", out)
+	}
+}

internal/lua/adapter_two_directions_test.go

@@ -0,0 +1,94 @@
+package lua
+
+import (
+	"encoding/json"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// TestEveryBundledAdapterTwoDirections 逐个验证每个内置适配器在**两个方向**上
+// 的声明与能力一致，且都能在真实 VM 上被调用而不报错。这是"重构后两个方向
+// 都没坏"的批量断言（用户要求：16 个适配器逐个验证两个方向）。
+//
+// 两个方向：
+//
+//	出向（egress）：serialize_* 钩子 —— 统一格式 → 客户端协议。
+//	   一个适配器要么声明了 serialize_*（接管），要么依赖内核 OpenAI 兜底。
+//	   两种都合法；不合法的是"声明了钩子但调用即报错"。
+//
+//	入向（ingress）：adapter.ingress + deserialize_request —— 客户端协议 → 统一格式。
+//	   声明 ingress 的适配器必须同时实现 deserialize_request，否则内核会让出给
+//	   一个空解析器（核心里 parseViaIngress 已挡住，此处再在适配器数据面钉一次）。
+//
+//	上游向（upstream）：transform_request —— 统一格式 → 上游协议。
+//	   这是所有适配器的本职，必须存在且可调用。
+func TestEveryBundledAdapterTwoDirections(t *testing.T) {
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	adapters := vm.ListAdapters()
+	if len(adapters) == 0 {
+		t.Fatal("no adapters loaded")
+	}
+
+	// 一个最小、协议中立的统一请求，供上游向 transform_request 使用。
+	unifiedReq := `{"model":"m","messages":[{"role":"user","content":"hi"}],"max_tokens":16,"stream":false}`
+
+	for _, a := range adapters {
+		a := a
+		t.Run(a.Name, func(t *testing.T) {
+			// ---- 上游向：transform_request 必须存在且可调用 ----
+			if !vm.HasFn(a.Name, "transform_request") {
+				t.Errorf("%s: 缺 transform_request（上游向本职）", a.Name)
+			} else if _, err := vm.Transform(a.Name, "transform_request", unifiedReq); err != nil {
+				t.Errorf("%s: transform_request 调用失败: %v", a.Name, err)
+			}
+
+			// ---- 出向：serialize_* 声明与调用一致性 ----
+			declaredEgress := false
+			for _, fn := range []string{"serialize_response", "serialize_chunk", "serialize_open_preamble", "serialize_close", "serialize_done_marker"} {
+				if vm.HasFn(a.Name, fn) {
+					declaredEgress = true
+				}
+			}
+			if vm.HasFn(a.Name, "serialize_response") {
+				// 统一结果形状（见 types.UnifiedResult）
+				res := `{"model":"m","id":"x","content":"hello","finish_reason":"stop","token_usage":{"prompt":1,"completion":1,"total":2}}`
+				out, err := vm.Transform(a.Name, "serialize_response", res)
+				if err != nil {
+					t.Errorf("%s: serialize_response 调用失败: %v", a.Name, err)
+				} else if strings.TrimSpace(out) == "" {
+					t.Errorf("%s: serialize_response 返回空", a.Name)
+				} else if !json.Valid([]byte(out)) && !strings.HasPrefix(strings.TrimSpace(out), "event:") {
+					t.Errorf("%s: serialize_response 既非 JSON 也非 SSE: %q", a.Name, out)
+				}
+			}
+			if declaredEgress {
+				t.Logf("%s: 出向由适配器接管", a.Name)
+			} else {
+				t.Logf("%s: 出向走内核 OpenAI 兜底（未声明 serialize_*）", a.Name)
+			}
+
+			// ---- 入向：声明 ingress 必须配 deserialize_request ----
+			ingress := vm.AdapterIngress(a.Name)
+			hasDeser := vm.HasFn(a.Name, "deserialize_request")
+			if ingress != "" {
+				if !hasDeser {
+					t.Errorf("%s: 声明 ingress=%q 但没有 deserialize_request —— 内核会让出给空解析器", a.Name, ingress)
+				} else {
+					// 用该适配器自己服务的那个端点的典型 body 调一次，确认不报错。
+					if _, err := vm.Transform(a.Name, "deserialize_request", `{"model":"m","max_tokens":8,"messages":[{"role":"user","content":"hi"}]}`); err != nil {
+						t.Errorf("%s: deserialize_request 调用失败: %v", a.Name, err)
+					}
+					t.Logf("%s: 入向认领 %s", a.Name, ingress)
+				}
+			} else if hasDeser {
+				t.Errorf("%s: 实现了 deserialize_request 却没声明 ingress —— 内核永远不会调用它", a.Name)
+			}
+		})
+	}
+}

internal/lua/adapter_validate.go

@@ -0,0 +1,133 @@
+// Package lua — 适配器导入安全校验（R6）
+//
+// 用户要求（feature/unified-wire-v2）：导入适配器时必须检查路径冲突，
+// 避免静默错误。
+//
+// 之前的风险：
+//
+//	UploadAdapter(name, code) 直接拼 adapter_dir/<name>.lua。
+//	name 含 "../" 或绝对路径会写到目录外；
+//	name 与内置适配器同名（openai/anthropic/gemini…）会静默覆盖——
+//	加载失败只报 "adapter X missing"，错误现场与病因分离。
+//
+// 这里的校验在**所有导入入口**统一执行（LoadAdapterSource / UploadAdapter /
+// 插件 install_adapter），让"为什么被拒绝"成为明确的错误信息。
+package lua
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// AdapterNamePattern 是适配器名的合法字符集。
+// 不允许：路径分隔符（/、\）、点点（..）、绝对路径、空格。
+// 允许：字母、数字、下划线、连字符。
+const AdapterNamePattern = `^[a-zA-Z0-9_-]+$`
+
+// ErrAdapterNameInvalid 适配器名不合法。
+type ErrAdapterNameInvalid struct{ Name, Reason string }
+
+func (e *ErrAdapterNameInvalid) Error() string {
+	return fmt.Sprintf("adapter name %q invalid: %s", e.Name, e.Reason)
+}
+
+// ErrAdapterNameReserved 适配器名与内核内置适配器冲突（用户要求：冲突时
+// 内核内置优先，外部不得覆盖）。
+type ErrAdapterNameReserved struct{ Name string }
+
+func (e *ErrAdapterNameReserved) Error() string {
+	return fmt.Sprintf("adapter name %q is reserved by a built-in adapter; renaming or deleting the built-in is not supported", e.Name)
+}
+
+// ValidateAdapterName 校验一个将被导入的适配器名。
+//
+// 它不做文件系统存在性检查（那是上传路径的责任），只校验"这个名字作为
+// adapter_dir/<name>.lua 的一部分是否安全"。校验维度：
+//
+//  1. 非空、符合字符集（拒绝路径穿越、绝对路径、空白）
+//  2. 不与内核内置适配器名冲突（除非 force 覆盖内置，但这通常不该被允许）
+//  3. 不含保留路径段（.、..、CON、NUL 等）
+//
+// reservedNames 是调用方提供的"保留名集合"——通常是内置适配器名表。
+// 传 nil 表示无保留名（仍校验字符集与路径安全）。
+func ValidateAdapterName(name string, reservedNames []string) error {
+	if name == "" {
+		return &ErrAdapterNameInvalid{Name: name, Reason: "empty"}
+	}
+	// 字符集：最简单的拒绝路径风险的方式。
+	if !matchesAdapterPattern(name) {
+		return &ErrAdapterNameInvalid{Name: name, Reason: "must match " + AdapterNamePattern + " (no slashes, dots-only, absolute paths, or whitespace)"}
+	}
+	// 保留段：即使字符集通过，也要点名拒绝 . 和 .. （它们会被某些字符集放行，
+	// 但绝不能成为文件名）。
+	switch name {
+	case ".", "..":
+		return &ErrAdapterNameInvalid{Name: name, Reason: "reserved path segment"}
+	}
+	for _, reserved := range reservedNames {
+		if strings.EqualFold(name, reserved) {
+			return &ErrAdapterNameReserved{Name: name}
+		}
+	}
+	return nil
+}
+
+// ValidateAdapterPath 在 LoadAdapter(path) 入口校验磁盘路径是否指向
+// adapter_dir 下、且文件名合法。
+//
+// 这一层是"从磁盘加载"的安全网：Start() 用它加载目录里的 .lua，必须保证
+// 一个被放进来的文件（无论谁放的）只能注册成它自己的名字，不能因为文件名
+// 怪异（如 ../openai.lua）就覆盖或越界。
+//
+// 返回校验后的 (name, error)。
+func ValidateAdapterPath(path, adapterDir string, reservedNames []string) (string, error) {
+	// 先解析成绝对路径，与 adapter_dir 比对，确认文件确实在目录内
+	// （不在目录内的 LoadAdapter 调用本就是异常用法，但 Start() 遍历的是
+	// ReadDir 的结果，天然在目录内，这条校验对它是 no-op）。
+	abs, err := filepath.Abs(path)
+	if err != nil {
+		return "", &ErrAdapterNameInvalid{Name: path, Reason: "unresolvable path"}
+	}
+	dirAbs, _ := filepath.Abs(adapterDir)
+	if !isWithinDir(abs, dirAbs) {
+		return "", &ErrAdapterNameInvalid{Name: path, Reason: "file is outside adapter_dir"}
+	}
+	name := strings.TrimSuffix(filepath.Base(path), filepath.Ext(path))
+	if err := ValidateAdapterName(name, reservedNames); err != nil {
+		return "", err
+	}
+	return name, nil
+}
+
+// matchesAdapterPattern 是 AdapterNamePattern 的纯函数实现（避免在 hot path
+// 里编译正则）。
+func matchesAdapterPattern(s string) bool {
+	if s == "" {
+		return false
+	}
+	for _, r := range s {
+		ok := (r >= 'a' && r <= 'z') ||
+			(r >= 'A' && r <= 'Z') ||
+			(r >= '0' && r <= '9') ||
+			r == '_' || r == '-'
+		if !ok {
+			return false
+		}
+	}
+	return true
+}
+
+// isWithinDir 判断 path 是否在 dir 内（含 path == dir 本身，不含 .. 逃逸）。
+func isWithinDir(path, dir string) bool {
+	rel, err := filepath.Rel(dir, path)
+	if err != nil {
+		return false
+	}
+	return rel != ".." && !strings.HasPrefix(rel, ".."+string(os.PathSeparator)) && rel != "."
+}
+
+// 哑引用：让 errors 包在 go vet 下不报"导出但未使用"。
+var _ = errors.New

internal/lua/adapter_validate_test.go

@@ -0,0 +1,197 @@
+package lua
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// 本文件锁定 R6：导入适配器时检查路径冲突，避免静默错误。
+//
+// 覆盖三类注入：
+//  1. 名字字符集（路径穿越 / 绝对路径 / 空白）→ 拒绝
+//  2. 保留名冲突（内置适配器）→ 拒绝
+//  3. 合法名 → 通过（且确实可加载）
+
+func TestValidateAdapterNameRejectsPathTraversal(t *testing.T) {
+	for _, name := range []string{
+		"../openai",   // 目录逃逸
+		"..",          // 保留段
+		".",           // 保留段
+		"/etc/passwd", // 绝对路径
+		"a/b/c",       // 斜杠
+		"a\\b",        // 反斜杠
+		"has space",   // 空白
+		"a b c",       // 空格
+		"",            // 空
+	} {
+		if err := ValidateAdapterName(name, nil); err == nil {
+			t.Errorf("名字 %q 应被拒绝（路径不安全），但通过了", name)
+		}
+	}
+}
+
+func TestValidateAdapterNameRejectsReserved(t *testing.T) {
+	reserved := []string{"openai", "anthropic", "gemini", "Ollama"}
+	for _, name := range []string{"openai", "anthropic", "gemini", "ollama"} {
+		if err := ValidateAdapterName(name, reserved); err == nil {
+			t.Errorf("名字 %q 与内置适配器冲突，应被拒绝", name)
+		}
+	}
+}
+
+func TestValidateAdapterNameAcceptsLegal(t *testing.T) {
+	for _, name := range []string{
+		"my_adapter",
+		"my-adapter",
+		"custom123",
+		"a",
+		"opencode", // 不是保留名（本次已从内置移除）
+	} {
+		if err := ValidateAdapterName(name, []string{"openai", "anthropic"}); err != nil {
+			t.Errorf("合法名字 %q 被误拒: %v", name, err)
+		}
+	}
+}
+
+// TestValidateAdapterPathRejectsOutsideDir 验证从磁盘加载路径时，
+// 目录外的文件被拒绝。
+func TestValidateAdapterPathRejectsOutsideDir(t *testing.T) {
+	dir := t.TempDir()
+	outside := filepath.Join(dir, "..", "evil.lua")
+	if _, err := ValidateAdapterPath(outside, dir, nil); err == nil {
+		t.Errorf("目录外的路径 %q 应被拒绝", outside)
+	}
+}
+
+// TestLoadAdapterSourceAllowsBuiltinOverride 验证运维覆盖内置适配器是允许的
+// （文档化的 adapter_dir 热覆盖机制）——仅插件被禁止影子内置。
+func TestLoadAdapterSourceAllowsBuiltinOverride(t *testing.T) {
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	code := `local a={} a.name="openai" a.version="override" a.endpoint="/x" return a`
+	if err := vm.LoadAdapterSource("openai", code); err != nil {
+		t.Errorf("运维覆盖内置 openai 应被允许（热覆盖机制），却被拒: %v", err)
+	}
+}
+
+// TestInstallPluginAdapterRejectsBuiltin 验证**插件**装内置名被拒（防篡改）。
+func TestInstallPluginAdapterRejectsBuiltin(t *testing.T) {
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	code := `local a={} a.name="openai" a.version="evil" a.endpoint="/x" return a`
+	if err := vm.InstallPluginAdapter("openai", code); err == nil {
+		t.Error("插件装与内置同名的适配器应被拒（防篡改发行物）")
+	}
+	// 但插件装非内置名应成功
+	ok := `local a={} a.name="myplugin_adapter" a.version="1" a.endpoint="/x" return a`
+	if err := vm.InstallPluginAdapter("myplugin_adapter", ok); err != nil {
+		t.Errorf("插件装非保留名应成功: %v", err)
+	}
+}
+
+func TestLoadAdapterSourceRejectsReserved(t *testing.T) {
+	t.Skip("已改：LoadAdapterSource 允许运维覆盖内置；保留名禁止已移至 InstallPluginAdapter")
+}
+
+// TestLoadAdapterSourceRejectsTraversal 验证加载入口拒绝路径穿越名。
+func TestLoadAdapterSourceRejectsTraversal(t *testing.T) {
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	code := `local a={} a.name="x" a.version="1" a.endpoint="/x" return a`
+	if err := vm.LoadAdapterSource("../evil", code); err == nil {
+		t.Error("路径穿越名 ../evil 应被拒绝")
+	}
+}
+
+// TestUploadAdapterPathTraversalDoesNotEscape 是端到端安全测试：
+// 模拟管理员通过 API 上传一个恶意名，验证它**不会**写到 adapter_dir 之外。
+func TestUploadAdapterPathTraversalDoesNotEscape(t *testing.T) {
+	// 用一个真实 Core 来测 UploadAdapter，但 core 需要完整配置——这里直接
+	// 测 UploadAdapter 的校验层（名字在写文件之前就被拒）。
+	// 由于 UploadAdapter 在 core 包，我们在 lua 层能测的等价面是
+	// ValidateAdapterName + LoadAdapterSource，上面已覆盖。
+	// 这里补一个"确无副作用"断言：遍历名不得创建任何文件。
+	dir := filepath.Join(t.TempDir(), "adapters")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := ValidateAdapterName("../openai", nil); err == nil {
+		t.Fatal("前提不成立")
+	}
+	// 证明没有 ../openai.lua 被创建
+	if _, err := os.Stat(filepath.Join(dir, "..", "openai.lua")); !os.IsNotExist(err) {
+		t.Error("校验前就写出了逃逸文件 —— 写入顺序错误")
+	}
+}
+
+// TestLoadAdapterValidatesTheDiskPath 验证从磁盘加载（Start 的路径）对路径
+// 做安全校验：目录外的文件被拒。（目录内同名内置的文件是合法覆盖，
+// 不在此拒绝——见 TestLoadAdapterSourceAllowsBuiltinOverride。）
+func TestLoadAdapterValidatesTheDiskPath(t *testing.T) {
+	dir := filepath.Join(t.TempDir(), "adapters")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	// 目录外的文件：应被拒
+	outside := filepath.Join(dir, "..", "evil.lua")
+	if err := os.WriteFile(outside, []byte(`local a={} a.name="evil" a.version="1" a.endpoint="/x" return a`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	vm := NewVM(dir)
+	if err := vm.LoadAdapter(outside); err == nil {
+		t.Error("目录外的磁盘文件应被拒绝")
+	}
+}
+
+// TestValidateAdapterPathInDirAccepted 验证目录内的合法文件通过。
+func TestValidateAdapterPathInDirAccepted(t *testing.T) {
+	dir := filepath.Join(t.TempDir(), "adapters")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	ok := filepath.Join(dir, "my_new.lua")
+	if err := os.WriteFile(ok, []byte(`local a={} a.name="my_new" a.version="1" a.endpoint="/x" return a`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	name, err := ValidateAdapterPath(ok, dir, []string{"openai"})
+	if err != nil {
+		t.Fatalf("合法文件被误拒: %v", err)
+	}
+	if name != "my_new" {
+		t.Errorf("导出名 = %q, want my_new", name)
+	}
+}
+
+// TestErrorNamesTheConflict 验证错误信息点名冲突对象（用户要求"避免静默
+// 错误"——错误必须可定位）。
+func TestErrorNamesTheConflict(t *testing.T) {
+	err := ValidateAdapterName("openai", []string{"openai"})
+	if err == nil {
+		t.Fatal("应拒绝")
+	}
+	var reserved *ErrAdapterNameReserved
+	if !errors.As(err, &reserved) {
+		t.Fatalf("错误类型应为 ErrAdapterNameReserved, got %T", err)
+	}
+	if !strings.Contains(err.Error(), "openai") {
+		t.Errorf("错误信息应点名冲突对象: %v", err)
+	}
+	if !strings.Contains(err.Error(), "built-in") {
+		t.Errorf("错误信息应说明冲突性质: %v", err)
+	}
+}

internal/lua/adapters/agentrouter.lua

@@ -1,212 +0,0 @@
-local adapter = {}
-
-adapter.name = "agentrouter"
-adapter.version = "1.0.0"
-adapter.endpoint = "/chat/completions"
-adapter.headers = {}
-
--- AgentRouter 的阿里云 WAF 按三重维度校验：出口 IP（必须海外）、
--- TLS 指纹（JS/Go HTTP 层与 curl 不同）、浏览器指纹 header 集。
--- 纯 UA 已不够：WAF 拦 12 种 agent CLI UA 中的 10 种（全 405 HTML），
--- 只有同时带 Origin: qwen.ai + Referer + X-Requested-With 才放行，
--- 且同 IP 短时间连续 3+ 次请求会被速率封禁，靠 adapter 侧低频调用 +
--- llmsproxy 的 prefFailStep 冷却自然限制，来源无法控制速率。
-local default_ua = "QwenCode/0.2.0 (linux; x64)"
-
--- AgentRouter fronts Claude models (claude-opus-4-8), and Claude upstreams
--- enforce Anthropic's tool id rule ^[a-zA-Z0-9_-]{1,64}$ by rejecting the WHOLE
--- request. Plain OpenAI does not, so an id minted by a permissive model (e.g.
--- "bash:0" from moonshotai/kimi-k3) is replayed here by the client and kills
--- every turn. See anthropic.lua / openai.lua for the same helper: Lua adapters
--- have no shared prelude, so each carries its own copy.
-local TOOL_ID_MAX = 64
-
-local function safe_tool_id(id)
-    if type(id) ~= "string" or id == "" then return id end
-    local clean = string.gsub(id, "[^A-Za-z0-9_-]", "_")
-    if clean == id and #clean <= TOOL_ID_MAX then
-        return clean
-    end
-    local digest = string.sub(sha256_hex(id), 1, 8)
-    local keep = TOOL_ID_MAX - #digest - 1
-    if #clean > keep then clean = string.sub(clean, 1, keep) end
-    return clean .. "_" .. digest
-end
-
-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
-    req.extra_body = nil
-    if req.messages then
-        for _, msg in ipairs(req.messages) do
-            msg.reasoning_content = nil
-            if msg.tool_call_id ~= nil then
-                msg.tool_call_id = safe_tool_id(msg.tool_call_id)
-            end
-            if type(msg.tool_calls) == "table" then
-                for _, tc in ipairs(msg.tool_calls) do
-                    if type(tc) == "table" and tc.id ~= nil then
-                        tc.id = safe_tool_id(tc.id)
-                    end
-                end
-            end
-        end
-    end
-    return json.encode(req)
-end
-
--- 注入 Authorization + 官方客户端 User-Agent
-function adapter.build_headers(meta)
-    local ua = default_ua
-    local meta_ua = (meta.source or {}).meta and (meta.source).meta.user_agent
-    if meta_ua and meta_ua ~= "" then
-        ua = meta_ua
-    end
-    return {
-        ["Content-Type"] = "application/json",
-        ["Authorization"] = "Bearer " .. meta.api_key,
-        ["User-Agent"] = ua,
-        -- QwenCode client fingerprint: WAF rejects bare UA without these.
-        ["Accept"] = "application/json, text/plain, */*",
-        ["Accept-Language"] = "zh-CN,zh;q=0.9",
-        ["Origin"] = "https://qwen.ai",
-        ["Referer"] = "https://qwen.ai/",
-        ["X-Requested-With"] = "XMLHttpRequest",
-        ["Connection"] = "keep-alive",
-    }
-end
-
-function adapter.transform_response(raw_body)
-    local ok, resp = pcall(json.decode, raw_body)
-    if not ok or resp == nil then return raw_body end
-
-    local unified = {
-        content = "",
-        finish_reason = "",
-        token_usage = { prompt = 0, completion = 0, total = 0 }
-    }
-
-    if type(resp.usage) == "table" then
-        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
-        unified.token_usage.completion = resp.usage.completion_tokens or 0
-        unified.token_usage.total = resp.usage.total_tokens or 0
-        local hit = 0
-        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
-            hit = resp.usage.prompt_tokens_details.cached_tokens
-            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
-        end
-        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
-            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
-            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
-            if hit == 0 then
-                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
-            end
-        end
-    end
-
-    if type(resp.choices) == "table" and #resp.choices > 0 then
-        local ch = resp.choices[1]
-        if type(ch.message) == "table" then
-            unified.content = ch.message.content or ""
-            if ch.message.reasoning_content then
-                unified.reasoning_content = ch.message.reasoning_content
-            end
-            if type(ch.message.tool_calls) == "table" then
-                local tcs = {}
-                for _, tc in ipairs(ch.message.tool_calls) do
-                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
-                    if not args_ok then args = {} end
-                    table.insert(tcs, {
-                        id = safe_tool_id(tc.id),
-                        type = tc.type or "function",
-                        name = tc["function"].name,
-                        arguments = args
-                    })
-                end
-                unified.tool_calls = tcs
-            end
-        end
-        unified.finish_reason = ch.finish_reason or ""
-    end
-
-    return json.encode(unified)
-end
-
-function adapter.transform_stream_chunk(raw_chunk)
-    local ok, chunk = pcall(json.decode, raw_chunk)
-    if not ok then return "" end
-
-    -- usage-only terminal chunk (empty choices + usage): forward it so the
-    -- gateway emits exact token counts (and cache fields when present).
-    local uses = nil
-    if type(chunk.usage) == "table" then
-        uses = {
-            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
-            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
-            total = chunk.usage.total_tokens or chunk.usage.total or 0,
-        }
-        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
-            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
-        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
-            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
-            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
-            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
-        end
-    end
-
-    if not chunk.choices or #chunk.choices == 0 then
-        if uses ~= nil then
-            return json.encode({ usage = uses, done = false })
-        end
-        return ""
-    end
-    local delta = chunk.choices[1].delta or {}
-    local fr = chunk.choices[1].finish_reason
-
-    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
-
-    local unified = {
-        content = delta.content or "",
-        done = (finish ~= nil)
-    }
-    if finish then
-        unified.finish_reason = finish
-    end
-    if delta.reasoning_content then
-        unified.reasoning_content = delta.reasoning_content
-    end
-    if delta.tool_calls then
-        -- Sanitize outbound too: a dirty id must never enter a client session,
-        -- because the client replays it to every other source. Only the first
-        -- fragment of a streamed call carries an id; later argument fragments
-        -- have none and must stay id-less for index-based accumulation.
-        for _, tc in ipairs(delta.tool_calls) do
-            if type(tc) == "table" and tc.id ~= nil then
-                tc.id = safe_tool_id(tc.id)
-            end
-        end
-        unified.tool_calls = delta.tool_calls
-    end
-    if uses ~= nil then
-        unified.usage = uses
-    end
-    return json.encode(unified)
-end
-
--- 错误收敛：前置 WAF 会返回整页 HTML（阿里云盾），JSON 时为 new-api 风格
-function adapter.transform_error(status, body)
-    local low = string.lower(body or "")
-    if string.sub(low, 1, 9) == "<!doctype" or string.find(low, "<html", 1, true) then
-        return "blocked by AgentRouter WAF"
-    end
-    local ok, resp = pcall(json.decode, body)
-    if not ok or type(resp) ~= "table" then return nil end
-    local e = resp.error
-    if type(e) == "table" and type(e.message) == "string" then
-        return e.message
-    end
-    return nil
-end
-
-return adapter

internal/lua/adapters/anthropic.lua

@@ -5,6 +5,14 @@ adapter.endpoint = "/v1/messages"
 adapter.headers = {
     ["anthropic-version"] = "2023-06-01",
 }
+-- INGRESS CLAIM (inbound dual of serialize_*): this adapter owns parsing of the
+-- client-facing POST /v1/messages request body. The kernel keeps a built-in
+-- Anthropic parser as its fallback (R5) — that is what an unconfigured
+-- deployment uses — but once this adapter is loaded the kernel YIELDS, and the
+-- deserialize_request hook below is the single place Anthropic's REQUEST shape
+-- is understood. Iterating on that shape therefore means editing this file,
+-- never the kernel.
+adapter.ingress = "/v1/messages"
 
 -- ============================================================
 -- helpers
@@ -132,6 +140,79 @@ local function append_user(msgs, blocks)
     end
 end
 
+-- ============================================================
+-- deserialize_request  (Anthropic → unified)   [INGRESS]
+-- ============================================================
+-- The inbound dual of serialize_response. The kernel hands over the raw client
+-- body and receives the unified request; from there scheduling is identical to
+-- /v1/chat/completions. Ported field-for-field from the kernel's former
+-- built-in anthropicRequest.toUnified — which remains as the kernel fallback
+-- for an unconfigured deployment.
+function adapter.deserialize_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok or type(req) ~= "table" then
+        error("invalid json")
+    end
+
+    -- system accepts Anthropic's two spellings: a plain string, or an array of
+    -- { type = "text", text = ... } blocks. Only text blocks contribute.
+    local system = ""
+    local sys = req.system
+    if type(sys) == "string" then
+        system = sys
+    elseif type(sys) == "table" then
+        for _, b in ipairs(sys) do
+            if type(b) == "table" and b.type == "text" and type(b.text) == "string" then
+                system = system .. b.text
+            end
+        end
+    end
+
+    local msgs = {}
+    local in_msgs = req.messages
+    if type(in_msgs) ~= "table" or #in_msgs == 0 then
+        error("messages is required")
+    end
+    for i, m in ipairs(in_msgs) do
+        if type(m) ~= "table" or type(m.role) ~= "string" or m.role == "" then
+            error("messages[" .. (i - 1) .. "].role is required")
+        end
+        -- Content is carried VERBATIM: a string stays a string, a block array
+        -- stays an array. The unified format keeps multimodal/tool blocks
+        -- untouched because only an adapter knows which block types the upstream
+        -- accepts. An absent/null content becomes the empty string.
+        local content = m.content
+        if content == nil or content == "" then
+            content = ""
+        end
+        table.insert(msgs, { role = m.role, content = content })
+    end
+
+    -- Anthropic's default is thinking OFF; emit that explicitly so the unified
+    -- request carries the client's intent rather than leaving it absent.
+    local thinking = req.thinking
+    if thinking == nil then
+        thinking = { enabled = false }
+    end
+
+    local out = {
+        model          = req.model or "",
+        messages       = msgs,
+        system         = system,
+        max_tokens     = req.max_tokens or 0,
+        stream         = req.stream == true,
+        temperature    = req.temperature,
+        top_p          = req.top_p,
+        top_k          = req.top_k or 0,
+        stop_sequences = req.stop_sequences,
+        tools          = req.tools,
+        tool_choice    = req.tool_choice,
+        thinking       = thinking,
+        extra_body     = {},
+    }
+    return json.encode(out)
+end
+
 -- ============================================================
 -- transform_request  (OpenAI → Anthropic)
 -- ============================================================
@@ -254,12 +335,24 @@ function adapter.transform_request(raw_body)
         end
     end
 
-    -- ── thinking (opt-in via extra_body) ──────────────────────
-    -- Default: OFF. Client sends extra_body.thinking to enable.
-    -- Example: {"extra_body": {"thinking": {"type": "enabled", "budget_tokens": 4096}}}
-    if type(req.extra_body) == "table" and type(req.extra_body.thinking) == "table" then
+    -- ── thinking ─────────────────────────────────────────────
+    -- Anthropic 的原生字段就是 {type, budget_tokens}，所以中立语义几乎直译。
+    --
+    -- 中立格式（v2）：{ enabled = true/false, budget_tokens = N }
+    --   enabled=false → 不传该字段（Anthropic 默认关闭思考，传 type=disabled
+    --                  是非法的，这也是过去只能「不传」的原因）
+    --   enabled=true  → { type="enabled", budget_tokens=N 或默认 }
+    --
+    -- 仍接受 extra_body.thinking 的原生写法：已有部署的配置不因此失效。
+    local th = req.thinking
+    if type(th) == "table" and th.enabled ~= false then
+        local budget = th.budget_tokens
+        if budget == nil or budget <= 0 then budget = 4096 end  -- Anthropic 必填
+        anthropic_req.thinking = { type = "enabled", budget_tokens = budget }
+    elseif type(req.extra_body) == "table" and type(req.extra_body.thinking) == "table" then
         anthropic_req.thinking = req.extra_body.thinking
     end
+    req.thinking = nil
 
     -- ── system ───────────────────────────────────────────────
     if system ~= "" then
@@ -444,4 +537,124 @@ function adapter.transform_error(status, body)
     return nil
 end
 
+-- ============================================================
+-- 对外序列化（统一格式 → Anthropic 形状）
+-- ============================================================
+--
+-- 内核只持统一格式；对外发什么形状由适配器决定（Phase 2）。
+-- 这几个钩子让以 anthropic 适配器为服务端的源，直接对 Anthropic 协议客户端
+-- 提供 /v1/messages 语义，而内核里不再有任何 Anthropic 字段名。
+--
+-- 与入向的 transform_* 互为逆运算：那里的映射（end_turn→stop）在这里反着
+-- 做回去（stop→end_turn），且默认 stop_sequence 是最保守的选择——它语义上
+-- 最接近"正常结束"。
+
+-- 统一 finish_reason → Anthropic stop_reason
+local function to_stop_reason(finish)
+    if finish == "tool_calls" then return "tool_use" end
+    if finish == "length" then return "max_tokens" end
+    if finish == "content_filter" then return "refusal" end
+    -- stop / 空 / 未知 → end_turn（Anthropic 的"正常说完"）
+    return "end_turn"
+end
+
+-- usage 数字化：内部 TokenUsage 的字段名与 Anthropic 的 input_tokens/output_tokens
+-- 不同名，这里做最后一次翻译。
+local function to_anthropic_usage(u)
+    if type(u) ~= "table" then
+        return { input_tokens = 0, output_tokens = 0 }
+    end
+    return {
+        input_tokens  = u.prompt or 0,
+        output_tokens = u.completion or 0,
+    }
+end
+
+-- 非流式响应：统一结果 → Anthropic Messages 响应对象
+function adapter.serialize_response(payload)
+    local p = json.decode(payload)
+    local blocks = {}
+    -- thinking 放最前，与 Anthropic 自己的块顺序一致
+    if p.reasoning ~= nil and p.reasoning ~= "" then
+        blocks[#blocks + 1] = { type = "thinking", thinking = p.reasoning }
+    end
+    blocks[#blocks + 1] = { type = "text", text = p.content or "" }
+    if type(p.tool_calls) == "table" and #p.tool_calls > 0 then
+        for _, tc in ipairs(p.tool_calls) do
+            blocks[#blocks + 1] = {
+                type = "tool_use",
+                id = tc.id or "",
+                name = tc.name or "",
+                input = tc.arguments or {},
+            }
+        end
+    end
+    return json.encode({
+        id = p.id or "",
+        type = "message",
+        role = "assistant",
+        model = p.model or "",
+        content = blocks,
+        stop_reason = to_stop_reason(p.finish_reason),
+        stop_sequence = nil,
+        usage = to_anthropic_usage(p.token_usage),
+    })
+end
+
+-- 角色前导：Anthropic 的流以 message_start 开头（不是 role delta）
+function adapter.serialize_open_preamble(payload)
+    local p = json.decode(payload)
+    return json.encode({
+        type = "message_start",
+        message = {
+            id = p.id or "",
+            type = "message",
+            role = "assistant",
+            model = p.model or "",
+            content = {},
+            stop_reason = nil,
+            stop_sequence = nil,
+            usage = { input_tokens = 0, output_tokens = 0 },
+        },
+    })
+end
+
+-- 流式增量：统一 chunk → Anthropic content_block_delta
+--
+-- 真实 Anthropic 流需要 content_block_start（开块）→ delta → stop（关块）
+-- 的三段式。这里用与多数 Anthropic 客户端兼容的简化形式：首个文本增量
+-- 带 index=0 的 delta，后续同 index；关闭由 serialize_close 负责。
+function adapter.serialize_chunk(payload)
+    local p = json.decode(payload)
+    if p.done then
+        -- 终止由 serialize_close 统一发，这里不重复
+        return ""
+    end
+    if p.reasoning ~= nil and p.reasoning ~= "" then
+        return json.encode({
+            type = "content_block_delta",
+            index = 0,
+            delta = { type = "thinking_delta", thinking = p.reasoning },
+        })
+    end
+    return json.encode({
+        type = "content_block_delta",
+        index = 0,
+        delta = { type = "text_delta", text = p.content or "" },
+    })
+end
+
+-- 收尾：message_delta（stop_reason + usage）→ message_stop
+function adapter.serialize_close(payload)
+    local p = json.decode(payload)
+    return json.encode({
+        type = "message_delta",
+        delta = {
+            stop_reason = to_stop_reason(p.finish_reason),
+            stop_sequence = nil,
+        },
+        usage = { output_tokens = 0 },
+    }) .. "\n" .. json.encode({ type = "message_stop" })
+end
+
 return adapter

internal/lua/adapters/deepseek.lua

@@ -15,11 +15,18 @@ function adapter.transform_request(raw_body)
         req.model = "deepseek-v4-flash"
     end
     req.stream = req.stream or false
-    if req.disable_thinking then
+    -- 中立的 thinking 语义 → DeepSeek 自己的协议写法。
+    -- 客户端说「关掉思考」时 thinking.enabled == false；上游的表达方式是
+    -- extra_body.thinking.type = "disabled"。v2 之前这里读的是内核私有字段
+    -- req.disable_thinking，而那个字段曾被内核泄漏进本函数收到的 body。
+    if type(req.thinking) == "table" and req.thinking.enabled == false then
         req.extra_body = req.extra_body or {}
         req.extra_body.thinking = { type = "disabled" }
+    elseif type(req.thinking) == "table" and req.thinking.budget_tokens ~= nil then
+        req.extra_body = req.extra_body or {}
+        req.extra_body.thinking = { type = "enabled", budget_tokens = req.thinking.budget_tokens }
     end
-    req.disable_thinking = nil
+    req.thinking = nil
 
     -- V4 thinking 模式要求：带 tool_calls 的 assistant 消息必须回传 reasoning_content。
     -- OpenAI 兼容客户端不会发该字段，补空串即可通过校验。

internal/lua/adapters/gemini.lua

@@ -18,6 +18,10 @@ adapter.version = "2.0.0"
 -- can never work. A source that overrides `endpoint:` bypasses the template
 -- entirely and must then spell the whole path itself.
 adapter.endpoint = "/v1beta/models/{model}:generateContent"
+-- 流式走同一路径的不同动词。这个事实以前写在 Go 内核里
+-- （provider.ChatURL 对 ":generateContent" 做字符串替换），属于把厂商协议
+-- 知识放进调度器；现在由适配器自己声明，内核只比较两个不透明字符串。
+adapter.endpoint_stream = "/v1beta/models/{model}:streamGenerateContent"
 adapter.headers = {}
 
 -- Gemini API: POST /v1/models/{model}:generateContent

internal/lua/adapters/github.lua

@@ -14,7 +14,13 @@ function adapter.transform_request(raw_body)
     req.temperature = req.temperature or 0.7
     req.max_tokens = req.max_tokens or 4096
     req.stream = req.stream or false
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/groq.lua

@@ -13,7 +13,13 @@ function adapter.transform_request(raw_body)
     req.temperature = req.temperature or 0.7
     req.max_tokens = req.max_tokens or 4096
     req.stream = req.stream or false
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/kimicode.lua

@@ -19,7 +19,13 @@ function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
     req.model = req.model or "kimi-k2"
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/mistral.lua

@@ -13,7 +13,13 @@ function adapter.transform_request(raw_body)
     req.temperature = req.temperature or 0.7
     req.max_tokens = req.max_tokens or 4096
     req.stream = req.stream or false
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/openai.lua

@@ -39,7 +39,13 @@ end
 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
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     if req.messages then
         for _, msg in ipairs(req.messages) do

internal/lua/adapters/opencode.lua

@@ -1,243 +0,0 @@
-local adapter = {}
-
-adapter.name = "opencode"
-adapter.version = "1.0.0"
-adapter.endpoint = "/chat/completions"
-
--- opencode.ai zen 网关按客户端指纹（UA + x-opencode-* 头）路由请求池：
--- 缺少 x-opencode-client/session/request/project 身份头的请求会被判为匿名
--- 客户端，x-preview-f-free 等模型在带 tools 时上游直接失败——流式返回单
--- chunk "finish_reason":"network_error" 空 content，非流式返回 503
--- "Endpoint is unavailable"，表现为"空回复"。因此除 UA 外必须附带
--- x-opencode-* 身份头；session/request 每请求派生唯一值（见 build_headers）。
-adapter.headers = {
-    ["User-Agent"] = "opencode/1.18.21 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14",
-}
-
--- 每请求生成身份头。沙箱无 os/math，用 meta.timestamp + 请求体哈希派生：
--- 同秒内重复请求 id 相同可接受（zen 只校验存在性，不校验格式）。
-local function rand_id(prefix, seed)
-    return prefix .. string.sub(sha256_hex(seed), 1, 24)
-end
-
-function adapter.build_headers(meta)
-    local ts = tostring(meta.timestamp or "")
-    return {
-        ["User-Agent"] = adapter.headers["User-Agent"],
-        ["x-opencode-client"] = "cli",
-        -- project 固定：同网关实例共享一个工作区身份
-        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. (meta.source and meta.source.name or "")), 1, 32),
-        ["x-opencode-session"] = rand_id("ses_", "session|" .. ts),
-        ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
-    }
-end
-
--- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
--- zen 上游 schema 只接受 text content part（无视觉/音频能力）：多模态 part
--- （image_url / input_audio / file 等）一律剥离。剥离后 content 变空的消息
--- 若不再携带 tool_calls / tool_call_id 才整条丢弃（避免上游
--- "unknown variant `image_url`, expected `text`"）；带工具调用的必须保留，
--- 否则会把紧随其后的 tool 结果变成孤儿，模型会反复重发同一个调用。
--- zen 上游角色白名单只有 system / user / assistant / tool / latest_reminder：
--- OpenAI 的 developer（及 function 等）不在其中，直接透传会触发上游
--- "unknown variant `developer`, expected one of ..." 错误；统一归一化为 system。
-local ROLE_WHITELIST = {
-    system = true,
-    user = true,
-    assistant = true,
-    tool = true,
-    latest_reminder = true,
-}
-
-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
-    req.extra_body = nil
-    -- stream_options is only valid alongside stream:true. Sending it on a
-    -- non-streaming request is rejected by stricter upstreams (OpenCode Go:
-    -- "stream_options should be set along with stream"), which failed every
-    -- non-stream call. Only attach it when the request actually streams.
-    if req.stream then
-        if type(req.stream_options) ~= "table" then req.stream_options = {} end
-        req.stream_options.include_usage = true
-    end
-    if req.messages then
-        local kept = {}
-        for _, msg in ipairs(req.messages) do
-            if type(msg.role) == "string" and not ROLE_WHITELIST[msg.role] then
-                msg.role = "system"
-            end
-            msg.reasoning_content = nil
-            local drop = false
-            if type(msg.content) == "table" then
-                local parts = {}
-                for _, part in ipairs(msg.content) do
-                    if type(part) == "table" and part.type ~= nil and part.type ~= "text" then
-                        -- multimodal part not supported by zen
-                    else
-                        table.insert(parts, part)
-                    end
-                end
-                if #parts == 0 then
-                    -- Content collapsed to nothing after stripping unsupported
-                    -- parts. A message that still carries a tool call must
-                    -- NEVER be dropped: the very next message is its tool
-                    -- result, and dropping the call orphans that result. The
-                    -- model then sees a result for a call it never made and
-                    -- re-issues the same tool call on every turn (observed as
-                    -- an infinite "repeated tool call" loop).
-                    --
-                    -- This is the common shape for agent clients: an assistant
-                    -- turn whose content is only [thinking, toolCall] serialises
-                    -- to content:[] with tool_calls — exactly the case that used
-                    -- to vanish here. Emit "" instead, which zen accepts.
-                    if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
-                        or msg.tool_call_id ~= nil then
-                        msg.content = ""
-                    else
-                        drop = true
-                    end
-                else
-                    msg.content = parts
-                end
-            end
-            if not drop then
-                table.insert(kept, msg)
-            end
-        end
-        req.messages = kept
-    end
-    return json.encode(req)
-end
-
-function adapter.transform_response(raw_body)
-    local ok, resp = pcall(json.decode, raw_body)
-    if not ok or resp == nil then return raw_body end
-
-    local unified = {
-        content = "",
-        finish_reason = "",
-        token_usage = { prompt = 0, completion = 0, total = 0 }
-    }
-
-    if type(resp.usage) == "table" then
-        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
-        unified.token_usage.completion = resp.usage.completion_tokens or 0
-        unified.token_usage.total = resp.usage.total_tokens or 0
-        local hit = 0
-        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
-            hit = resp.usage.prompt_tokens_details.cached_tokens
-            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
-        end
-        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
-            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
-            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
-            if hit == 0 then
-                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
-            end
-        end
-    end
-
-    if type(resp.choices) == "table" and #resp.choices > 0 then
-        local ch = resp.choices[1]
-        if type(ch.message) == "table" then
-            unified.content = ch.message.content or ""
-            local reasoning = ch.message.reasoning_content or ch.message.reasoning
-            if reasoning then
-                unified.reasoning_content = reasoning
-            end
-            if type(ch.message.tool_calls) == "table" then
-                local tcs = {}
-                for _, tc in ipairs(ch.message.tool_calls) do
-                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
-                    if not args_ok then args = {} end
-                    table.insert(tcs, {
-                        id = tc.id,
-                        type = tc.type or "function",
-                        name = tc["function"].name,
-                        arguments = args
-                    })
-                end
-                unified.tool_calls = tcs
-            end
-        end
-        unified.finish_reason = ch.finish_reason or ""
-    end
-
-    return json.encode(unified)
-end
-
-function adapter.transform_stream_chunk(raw_chunk)
-    local ok, chunk = pcall(json.decode, raw_chunk)
-    if not ok then return "" end
-
-    -- OpenAI-style streams may attach usage to a chunk with empty choices
-    -- (the final usage chunk). Preserve it; the gateway emits it as the
-    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
-    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
-    local uses = nil
-    if type(chunk.usage) == "table" then
-        uses = {
-            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
-            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
-            total = chunk.usage.total_tokens or chunk.usage.total or 0,
-        }
-        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
-            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
-        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
-            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
-            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
-            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
-        end
-    end
-
-    if not chunk.choices or #chunk.choices == 0 then
-        if uses ~= nil then
-            -- usage-only chunk is not a content/finish signal; the gateway
-            -- emits its own terminal stop chunk and merges this usage.
-            return json.encode({ usage = uses, done = false })
-        end
-        return ""
-    end
-    local delta = chunk.choices[1].delta or {}
-    local fr = chunk.choices[1].finish_reason
-
-    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
-
-    local unified = {
-        content = delta.content or "",
-        done = (finish ~= nil)
-    }
-    if finish then
-        unified.finish_reason = finish
-    end
-    if uses ~= nil then
-        unified.usage = uses
-    end
-    -- zen 用 reasoning 字段承载推理文本（OpenAI 惯例是 reasoning_content）
-    local reasoning = delta.reasoning_content or delta.reasoning
-    if reasoning then
-        unified.reasoning_content = reasoning
-    end
-    if delta.tool_calls then
-        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
-        unified.tool_calls = delta.tool_calls
-    end
-    return json.encode(unified)
-end
-
--- 错误收敛（可选钩子）：zen 错误信封固定为 {error={type,message}}；
--- 免费池限流 FreeUsageLimitError 单独标注。返回 nil 走通用兜底。
-function adapter.transform_error(status, body)
-    local ok, resp = pcall(json.decode, body)
-    if not ok or type(resp) ~= "table" then return nil end
-    local e = resp.error
-    if type(e) ~= "table" then return nil end
-    if e.type == "FreeUsageLimitError" then
-        return "zen free pool quota exhausted"
-    end
-    return e.message
-end
-
-return adapter

internal/lua/adapters/opencodego.lua

@@ -1,354 +0,0 @@
-local adapter = {}
-
-adapter.name = "opencodego"
-adapter.version = "1.0.0"
-adapter.endpoint = "/chat/completions"
-
--- OpenCode Go（包月订阅端点，https://opencode.ai/zen/go/v1）
---
--- 与 OpenCode Zen（opencodezen.lua，https://opencode.ai/zen/v1）是两个不同的
--- 服务，行为要求并不相同，因此各有专用适配器：
---
---   * 本适配器（Go）服务包月订阅池。上游**强制要求**
---     `x-opencode-session` 头，缺失直接 400 "MissingSessionID"。
---   * Go 的 thinking 模型（deepseek-v4.1-flash 等）**必须回传** assistant 轮的
---     `reasoning_content`，否则 400：
---       The `reasoning_content` in the thinking mode must be passed back to the API.
---     这是本适配器与 Zen 适配器最关键的差异 —— Zen 会剥掉它，Go 必须原样保留。
---   * Go 上游对协议更严格：非流式请求带 `stream_options` 会被拒
---     （"stream_options should be set along with stream"）。
---
--- 同样需要 opencode 客户端指纹（UA + x-opencode-*）才能被正确识别与路由。
-adapter.headers = {
-    ["User-Agent"] = "opencode/1.18.21 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14",
-}
-
--- 每请求生成身份头。沙箱无 os/math，用 meta.timestamp + 请求体哈希派生：
--- 同秒内重复请求 id 相同可接受（zen 只校验存在性，不校验格式）。
-local function rand_id(prefix, seed)
-    return prefix .. string.sub(sha256_hex(seed), 1, 24)
-end
-
--- 会话指纹：取历史里**第一条 user 消息**。
---
--- 为什么不直接用客户端的会话 id：实测抓包（tcpdump 抓 127.0.0.1:8081 的真实
--- agent 请求）确认通用 OpenAI 客户端**根本不发**任何会话标识 —— body 里没有
--- user / session_id / conversation_id / metadata，请求头也只有 X-Stainless-*
--- （OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套 x-opencode-session
--- 是它自己的概念，通用客户端无从转发。
---
--- 退而求其次但要够用：历史被逐轮重放，**第一条 user 消息在整个会话里恒定**，
--- 用它做指纹即可得到"每会话一个 session"，而不是"每源一个 session"。
--- 拿不到时（无 user 消息）返回空串，调用方回退到按源稳定。
-local function conversation_fingerprint(body)
-    if type(body) ~= "string" or body == "" then return "" end
-    local ok, req = pcall(json.decode, body)
-    if not ok or type(req) ~= "table" then return "" end
-    for _, m in ipairs(req.messages or {}) do
-        if type(m) == "table" and m.role == "user" then
-            local c = m.content
-            if type(c) == "string" then
-                if c ~= "" then return c end
-            elseif type(c) == "table" then
-                for _, part in ipairs(c) do
-                    if type(part) == "table" and part.type == "text" and part.text then
-                        return part.text
-                    end
-                end
-            end
-            return ""
-        end
-    end
-    return ""
-end
-
--- session_seed 决定上游会话号（x-opencode-session 的种子）。
---
--- 优先级：
---   1) 客户端自带的会话标识（meta.client_session）——pi 等在开启
---      compat.sendSessionAffinityHeaders 后会发 x-session-affinity，
---      这是平台真实的会话 id，整个会话恒定且天然按会话隔离。
---   2) 退路：历史里**第一条 user 消息**做会话指纹。
---   3) 再退：只按源固定（连 user 消息都没有时）。
---
--- 为什么需要 2/3：通用 OpenAI 客户端默认**根本不发**会话标识 —— 实测抓包
--- （tcpdump 抓 127.0.0.1:8081 真实 agent 请求）确认 body 里没有
--- user / session_id / conversation_id / metadata，请求头也只有
--- X-Stainless-*（OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套
--- x-opencode-session 是它自己的概念，通用客户端无从转发。
---
--- 为什么必须稳定：上游前缀缓存是**会话级**的。会话号每请求一变，
--- 缓存永不命中（实测：固定会号第 2 次命中 5888，每请求换会号则恒为 0）。
-local function session_seed(meta)
-    local src = (meta.source and meta.source.name) or ""
-    local base = "session|llmsproxy|" .. src
-    local cs = meta.client_session
-    if type(cs) == "string" and cs ~= "" then
-        return base .. "|client|" .. cs
-    end
-    return base .. "|" .. conversation_fingerprint(meta.body)
-end
-
-function adapter.build_headers(meta)
-    local ts = tostring(meta.timestamp or "")
-    local src = (meta.source and meta.source.name) or ""
-    return {
-        ["User-Agent"] = adapter.headers["User-Agent"],
-        ["x-opencode-client"] = "cli",
-        -- project 固定：同网关实例共享一个工作区身份
-        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. src), 1, 32),
-        -- session 必须**按源稳定**，不能每请求换：上游的前缀缓存在同一 session
-        -- 内才复用。实测（同一段 6032 token 提示词）：
-        --   固定 session  -> 第 2 次命中 5888/6032，cached_tokens=5888
-        --   每请求换 session -> 永远 0 命中
-        -- 原先用 meta.timestamp 派生，等于每请求都是新会话，缓存永远无效，
-        -- 上游也无法做会话亲和路由。
-        ["x-opencode-session"] = rand_id("ses_", session_seed(meta)),
-        -- request id 仍每请求唯一（它只是请求标识，不参与缓存键）
-        ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
-    }
-end
-
--- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
--- zen 上游 schema 只接受 text content part（无视觉/音频能力）：多模态 part
--- （image_url / input_audio / file 等）一律剥离。剥离后 content 变空的消息
--- 若不再携带 tool_calls / tool_call_id 才整条丢弃（避免上游
--- "unknown variant `image_url`, expected `text`"）；带工具调用的必须保留，
--- 否则会把紧随其后的 tool 结果变成孤儿，模型会反复重发同一个调用。
--- zen 上游角色白名单只有 system / user / assistant / tool / latest_reminder：
--- OpenAI 的 developer（及 function 等）不在其中，直接透传会触发上游
--- "unknown variant `developer`, expected one of ..." 错误；统一归一化为 system。
-local ROLE_WHITELIST = {
-    system = true,
-    user = true,
-    assistant = true,
-    tool = true,
-    latest_reminder = true,
-}
-
-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
-    req.extra_body = nil
-    -- stream_options is only valid alongside stream:true; sending it on a
-    -- non-streaming request is rejected by strict upstreams.
-    if req.stream then
-        if type(req.stream_options) ~= "table" then req.stream_options = {} end
-        req.stream_options.include_usage = true
-    end
-    if req.messages then
-        local kept = {}
-        for _, msg in ipairs(req.messages) do
-            if type(msg.role) == "string" and not ROLE_WHITELIST[msg.role] then
-                msg.role = "system"
-            end
-            -- Go 的 thinking 模式**必须**回传 reasoning_content：剥掉它会让
-            -- 上游直接 400（"The `reasoning_content` in the thinking mode must
-            -- be passed back to the API"），agent 的每一轮都会失败。
-            -- 注意：与 opencodezen.lua 的行为**相反**，不要在这里剥。
-            --
-            -- 但绝大多数客户端（pi 等）根本不保存也不回传 reasoning，只保留
-            -- 工具调用本身。上游只在“带 tool_calls 的助手轮”上校验这个字段，
-            -- 实测**空串即可通过校验**，所以缺省时补空串：既满足上游的
-            -- 一致性要求，又不伪造任何推理内容（用户看到的 reasoning 仍然是
-            -- 上游本轮真实返回的）。
-            if msg.role == "assistant"
-                and type(msg.tool_calls) == "table" and #msg.tool_calls > 0
-                and msg.reasoning_content == nil then
-                msg.reasoning_content = ""
-            end
-            local drop = false
-            if type(msg.content) == "table" then
-                local parts = {}
-                for _, part in ipairs(msg.content) do
-                    if type(part) == "table" and part.type ~= nil and part.type ~= "text" then
-                        -- multimodal part not supported by zen
-                    else
-                        table.insert(parts, part)
-                    end
-                end
-                if #parts == 0 then
-                    -- Content collapsed to nothing after stripping unsupported
-                    -- parts. A message that still carries a tool call must
-                    -- NEVER be dropped: the very next message is its tool
-                    -- result, and dropping the call orphans that result. The
-                    -- model then sees a result for a call it never made and
-                    -- re-issues the same tool call on every turn (observed as
-                    -- an infinite "repeated tool call" loop).
-                    if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
-                        or msg.tool_call_id ~= nil then
-                        msg.content = ""
-                    else
-                        drop = true
-                    end
-                else
-                    msg.content = parts
-                end
-            end
-            if not drop then
-                table.insert(kept, msg)
-            end
-        end
-        req.messages = kept
-    end
-    return json.encode(req)
-end
-
-function adapter.transform_response(raw_body)
-    local ok, resp = pcall(json.decode, raw_body)
-    if not ok or resp == nil then return raw_body end
-
-    local unified = {
-        content = "",
-        finish_reason = "",
-        token_usage = { prompt = 0, completion = 0, total = 0 }
-    }
-
-    if type(resp.usage) == "table" then
-        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
-        unified.token_usage.completion = resp.usage.completion_tokens or 0
-        unified.token_usage.total = resp.usage.total_tokens or 0
-        local hit = 0
-        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
-            hit = resp.usage.prompt_tokens_details.cached_tokens
-            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
-        end
-        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
-            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
-            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
-            if hit == 0 then
-                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
-            end
-        end
-        -- 上游报告输出里有多少是思考 token。不给客户端的话，无法判断
-        -- completion_tokens 里多少是「可见回答」、多少是「思考」（都按输出计费）。
-        local ctd = resp.usage.completion_tokens_details
-        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
-            unified.token_usage.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
-        end
-    end
-
-    -- 本次调用的费用，上游放在**顶层**且是**字符串**（如 "0"、"0.0012"）。
-    -- Go 订阅是包月制，恒为 "0"；只有 Zen 按量付费模型才有信息量。
-    -- 原样透传：网关不解析、不换算、不汇总 —— 它只是上游事实的搬运者。
-    if resp.cost ~= nil then
-        unified.cost = tostring(resp.cost)
-    end
-
-    if type(resp.choices) == "table" and #resp.choices > 0 then
-        local ch = resp.choices[1]
-        if type(ch.message) == "table" then
-            unified.content = ch.message.content or ""
-            local reasoning = ch.message.reasoning_content or ch.message.reasoning
-            if reasoning then
-                unified.reasoning_content = reasoning
-            end
-            if type(ch.message.tool_calls) == "table" then
-                local tcs = {}
-                for _, tc in ipairs(ch.message.tool_calls) do
-                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
-                    if not args_ok then args = {} end
-                    table.insert(tcs, {
-                        id = tc.id,
-                        type = tc.type or "function",
-                        name = tc["function"].name,
-                        arguments = args
-                    })
-                end
-                unified.tool_calls = tcs
-            end
-        end
-        unified.finish_reason = ch.finish_reason or ""
-    end
-
-    return json.encode(unified)
-end
-
-function adapter.transform_stream_chunk(raw_chunk)
-    local ok, chunk = pcall(json.decode, raw_chunk)
-    if not ok then return "" end
-
-    -- OpenAI-style streams may attach usage to a chunk with empty choices
-    -- (the final usage chunk). Preserve it; the gateway emits it as the
-    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
-    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
-    local uses = nil
-    if type(chunk.usage) == "table" then
-        uses = {
-            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
-            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
-            total = chunk.usage.total_tokens or chunk.usage.total or 0,
-        }
-        -- 独立字段与 prompt_tokens_details **并存**透传，不要写成 elseif：
-        -- 上游命中缓存时两者都发，非流式路径也是两个都带。写成 elseif 会让
-        -- 流式丢掉 prompt_cache_hit_tokens/miss，与同一源的**非流式**产出不一致
-        -- （dsh 优先读 details，但只认独立字段的老客户端会看不到缓存）。
-        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
-            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
-        end
-        if (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
-            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
-            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
-            if uses.prompt_tokens_details == nil then
-                uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
-            end
-        end
-        local ctd = chunk.usage.completion_tokens_details
-        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
-            uses.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
-        end
-    end
-
-    if not chunk.choices or #chunk.choices == 0 then
-        -- 空 choices 的帧既不是内容也不是结束信号。上游发两种：
-        --   {"choices":[],"usage":{...}}   —— 终帧用量
-        --   {"choices":[],"cost":"0"}      —— 单独的费用帧
-        -- 两者可能落在同一帧上，必须都转出去；早退只带 usage 会把费用丢干净。
-        local out = { done = false }
-        if uses ~= nil then out.usage = uses end
-        if chunk.cost ~= nil then out.cost = tostring(chunk.cost) end
-        if uses == nil and chunk.cost == nil then return "" end
-        return json.encode(out)
-    end
-    local delta = chunk.choices[1].delta or {}
-    local fr = chunk.choices[1].finish_reason
-
-    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
-
-    local unified = {
-        content = delta.content or "",
-        done = (finish ~= nil)
-    }
-    if finish then
-        unified.finish_reason = finish
-    end
-    if uses ~= nil then
-        unified.usage = uses
-    end
-    -- zen 用 reasoning 字段承载推理文本（OpenAI 惯例是 reasoning_content）
-    local reasoning = delta.reasoning_content or delta.reasoning
-    if reasoning then
-        unified.reasoning_content = reasoning
-    end
-    if delta.tool_calls then
-        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
-        unified.tool_calls = delta.tool_calls
-    end
-    return json.encode(unified)
-end
-
--- 错误收敛（可选钩子）：zen 错误信封固定为 {error={type,message}}；
--- 免费池限流 FreeUsageLimitError 单独标注。返回 nil 走通用兜底。
-function adapter.transform_error(status, body)
-    local ok, resp = pcall(json.decode, body)
-    if not ok or type(resp) ~= "table" then return nil end
-    local e = resp.error
-    if type(e) ~= "table" then return nil end
-    if e.type == "FreeUsageLimitError" then
-        return "zen free pool quota exhausted"
-    end
-    return e.message
-end
-
-return adapter

internal/lua/adapters/opencodezen.lua

@@ -1,342 +0,0 @@
-local adapter = {}
-
-adapter.name = "opencodezen"
-adapter.version = "1.0.0"
-adapter.endpoint = "/chat/completions"
-
--- OpenCode Zen（按量付费端点，https://opencode.ai/zen/v1）
---
--- 与 OpenCode Go（opencodego.lua，https://opencode.ai/zen/go/v1）是两个不同的
--- 服务，行为要求并不相同，因此各有专用适配器：
---
---   * 本适配器（Zen）服务免费池 / 按量付费池。免费模型必须带 opencode 客户端
---     指纹（UA + x-opencode-*），否则上游拒绝："free tier can only be used in
---     OpenCode"。Zen 的免费池以非 thinking 模型为主，历史上回传
---     reasoning_content 会引出上游报错，故这里剥掉（Go 侧相反，见 opencodego.lua）。
---   * Go 订阅读者请看 opencodego.lua：那边**必须**回传 reasoning_content，
---     否则 thinking 模式直接 400。
---
--- 缺 x-opencode-client/session/request/project 身份头会被判为匿名客户端，
--- 部分模型在带 tools 时上游直接失败（流式返回单 chunk
--- "finish_reason":"network_error" 空 content，非流式 503
--- "Endpoint is unavailable"，表现为"空回复"）。
-adapter.headers = {
-    ["User-Agent"] = "opencode/1.18.21 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14",
-}
-
--- 每请求生成身份头。沙箱无 os/math，用 meta.timestamp + 请求体哈希派生：
--- 同秒内重复请求 id 相同可接受（zen 只校验存在性，不校验格式）。
-local function rand_id(prefix, seed)
-    return prefix .. string.sub(sha256_hex(seed), 1, 24)
-end
-
--- 会话指纹：取历史里**第一条 user 消息**。
---
--- 为什么不直接用客户端的会话 id：实测抓包（tcpdump 抓 127.0.0.1:8081 的真实
--- agent 请求）确认通用 OpenAI 客户端**根本不发**任何会话标识 —— body 里没有
--- user / session_id / conversation_id / metadata，请求头也只有 X-Stainless-*
--- （OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套 x-opencode-session
--- 是它自己的概念，通用客户端无从转发。
---
--- 退而求其次但要够用：历史被逐轮重放，**第一条 user 消息在整个会话里恒定**，
--- 用它做指纹即可得到"每会话一个 session"，而不是"每源一个 session"。
--- 拿不到时（无 user 消息）返回空串，调用方回退到按源稳定。
-local function conversation_fingerprint(body)
-    if type(body) ~= "string" or body == "" then return "" end
-    local ok, req = pcall(json.decode, body)
-    if not ok or type(req) ~= "table" then return "" end
-    for _, m in ipairs(req.messages or {}) do
-        if type(m) == "table" and m.role == "user" then
-            local c = m.content
-            if type(c) == "string" then
-                if c ~= "" then return c end
-            elseif type(c) == "table" then
-                for _, part in ipairs(c) do
-                    if type(part) == "table" and part.type == "text" and part.text then
-                        return part.text
-                    end
-                end
-            end
-            return ""
-        end
-    end
-    return ""
-end
-
--- session_seed 决定上游会话号（x-opencode-session 的种子）。
---
--- 优先级：
---   1) 客户端自带的会话标识（meta.client_session）——pi 等在开启
---      compat.sendSessionAffinityHeaders 后会发 x-session-affinity，
---      这是平台真实的会话 id，整个会话恒定且天然按会话隔离。
---   2) 退路：历史里**第一条 user 消息**做会话指纹。
---   3) 再退：只按源固定（连 user 消息都没有时）。
---
--- 为什么需要 2/3：通用 OpenAI 客户端默认**根本不发**会话标识 —— 实测抓包
--- （tcpdump 抓 127.0.0.1:8081 真实 agent 请求）确认 body 里没有
--- user / session_id / conversation_id / metadata，请求头也只有
--- X-Stainless-*（OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套
--- x-opencode-session 是它自己的概念，通用客户端无从转发。
---
--- 为什么必须稳定：上游前缀缓存是**会话级**的。会话号每请求一变，
--- 缓存永不命中（实测：固定会号第 2 次命中 5888，每请求换会号则恒为 0）。
-local function session_seed(meta)
-    local src = (meta.source and meta.source.name) or ""
-    local base = "session|llmsproxy|" .. src
-    local cs = meta.client_session
-    if type(cs) == "string" and cs ~= "" then
-        return base .. "|client|" .. cs
-    end
-    return base .. "|" .. conversation_fingerprint(meta.body)
-end
-
-function adapter.build_headers(meta)
-    local ts = tostring(meta.timestamp or "")
-    local src = (meta.source and meta.source.name) or ""
-    return {
-        ["User-Agent"] = adapter.headers["User-Agent"],
-        ["x-opencode-client"] = "cli",
-        -- project 固定：同网关实例共享一个工作区身份
-        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. src), 1, 32),
-        -- session 必须**按源稳定**，不能每请求换：上游的前缀缓存在同一 session
-        -- 内才复用。实测（同一段 6032 token 提示词）：
-        --   固定 session  -> 第 2 次命中 5888/6032，cached_tokens=5888
-        --   每请求换 session -> 永远 0 命中
-        -- 原先用 meta.timestamp 派生，等于每请求都是新会话，缓存永远无效，
-        -- 上游也无法做会话亲和路由。
-        ["x-opencode-session"] = rand_id("ses_", session_seed(meta)),
-        -- request id 仍每请求唯一（它只是请求标识，不参与缓存键）
-        ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
-    }
-end
-
--- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
--- zen 上游 schema 只接受 text content part（无视觉/音频能力）：多模态 part
--- （image_url / input_audio / file 等）一律剥离。剥离后 content 变空的消息
--- 若不再携带 tool_calls / tool_call_id 才整条丢弃（避免上游
--- "unknown variant `image_url`, expected `text`"）；带工具调用的必须保留，
--- 否则会把紧随其后的 tool 结果变成孤儿，模型会反复重发同一个调用。
--- zen 上游角色白名单只有 system / user / assistant / tool / latest_reminder：
--- OpenAI 的 developer（及 function 等）不在其中，直接透传会触发上游
--- "unknown variant `developer`, expected one of ..." 错误；统一归一化为 system。
-local ROLE_WHITELIST = {
-    system = true,
-    user = true,
-    assistant = true,
-    tool = true,
-    latest_reminder = true,
-}
-
-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
-    req.extra_body = nil
-    -- stream_options is only valid alongside stream:true; sending it on a
-    -- non-streaming request is rejected by strict upstreams.
-    if req.stream then
-        if type(req.stream_options) ~= "table" then req.stream_options = {} end
-        req.stream_options.include_usage = true
-    end
-    if req.messages then
-        local kept = {}
-        for _, msg in ipairs(req.messages) do
-            if type(msg.role) == "string" and not ROLE_WHITELIST[msg.role] then
-                msg.role = "system"
-            end
-            -- Zen 免费池不接受回传 reasoning_content（Go 侧相反）
-            msg.reasoning_content = nil
-            local drop = false
-            if type(msg.content) == "table" then
-                local parts = {}
-                for _, part in ipairs(msg.content) do
-                    if type(part) == "table" and part.type ~= nil and part.type ~= "text" then
-                        -- multimodal part not supported by zen
-                    else
-                        table.insert(parts, part)
-                    end
-                end
-                if #parts == 0 then
-                    -- Content collapsed to nothing after stripping unsupported
-                    -- parts. A message that still carries a tool call must
-                    -- NEVER be dropped: the very next message is its tool
-                    -- result, and dropping the call orphans that result. The
-                    -- model then sees a result for a call it never made and
-                    -- re-issues the same tool call on every turn (observed as
-                    -- an infinite "repeated tool call" loop).
-                    if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
-                        or msg.tool_call_id ~= nil then
-                        msg.content = ""
-                    else
-                        drop = true
-                    end
-                else
-                    msg.content = parts
-                end
-            end
-            if not drop then
-                table.insert(kept, msg)
-            end
-        end
-        req.messages = kept
-    end
-    return json.encode(req)
-end
-
-function adapter.transform_response(raw_body)
-    local ok, resp = pcall(json.decode, raw_body)
-    if not ok or resp == nil then return raw_body end
-
-    local unified = {
-        content = "",
-        finish_reason = "",
-        token_usage = { prompt = 0, completion = 0, total = 0 }
-    }
-
-    if type(resp.usage) == "table" then
-        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
-        unified.token_usage.completion = resp.usage.completion_tokens or 0
-        unified.token_usage.total = resp.usage.total_tokens or 0
-        local hit = 0
-        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
-            hit = resp.usage.prompt_tokens_details.cached_tokens
-            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
-        end
-        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
-            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
-            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
-            if hit == 0 then
-                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
-            end
-        end
-        -- 上游报告输出里有多少是思考 token。不给客户端的话，无法判断
-        -- completion_tokens 里多少是「可见回答」、多少是「思考」（都按输出计费）。
-        local ctd = resp.usage.completion_tokens_details
-        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
-            unified.token_usage.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
-        end
-    end
-
-    -- 本次调用的费用，上游放在**顶层**且是**字符串**（如 "0"、"0.0012"）。
-    -- Go 订阅是包月制，恒为 "0"；只有 Zen 按量付费模型才有信息量。
-    -- 原样透传：网关不解析、不换算、不汇总 —— 它只是上游事实的搬运者。
-    if resp.cost ~= nil then
-        unified.cost = tostring(resp.cost)
-    end
-
-    if type(resp.choices) == "table" and #resp.choices > 0 then
-        local ch = resp.choices[1]
-        if type(ch.message) == "table" then
-            unified.content = ch.message.content or ""
-            local reasoning = ch.message.reasoning_content or ch.message.reasoning
-            if reasoning then
-                unified.reasoning_content = reasoning
-            end
-            if type(ch.message.tool_calls) == "table" then
-                local tcs = {}
-                for _, tc in ipairs(ch.message.tool_calls) do
-                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
-                    if not args_ok then args = {} end
-                    table.insert(tcs, {
-                        id = tc.id,
-                        type = tc.type or "function",
-                        name = tc["function"].name,
-                        arguments = args
-                    })
-                end
-                unified.tool_calls = tcs
-            end
-        end
-        unified.finish_reason = ch.finish_reason or ""
-    end
-
-    return json.encode(unified)
-end
-
-function adapter.transform_stream_chunk(raw_chunk)
-    local ok, chunk = pcall(json.decode, raw_chunk)
-    if not ok then return "" end
-
-    -- OpenAI-style streams may attach usage to a chunk with empty choices
-    -- (the final usage chunk). Preserve it; the gateway emits it as the
-    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
-    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
-    local uses = nil
-    if type(chunk.usage) == "table" then
-        uses = {
-            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
-            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
-            total = chunk.usage.total_tokens or chunk.usage.total or 0,
-        }
-        -- 独立字段与 prompt_tokens_details **并存**透传，不要写成 elseif：
-        -- 上游命中缓存时两者都发，非流式路径也是两个都带。写成 elseif 会让
-        -- 流式丢掉 prompt_cache_hit_tokens/miss，与同一源的**非流式**产出不一致
-        -- （dsh 优先读 details，但只认独立字段的老客户端会看不到缓存）。
-        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
-            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
-        end
-        if (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
-            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
-            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
-            if uses.prompt_tokens_details == nil then
-                uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
-            end
-        end
-        local ctd = chunk.usage.completion_tokens_details
-        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
-            uses.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
-        end
-    end
-
-    if not chunk.choices or #chunk.choices == 0 then
-        -- 空 choices 的帧既不是内容也不是结束信号。上游发两种：
-        --   {"choices":[],"usage":{...}}   —— 终帧用量
-        --   {"choices":[],"cost":"0"}      —— 单独的费用帧
-        -- 两者可能落在同一帧上，必须都转出去；早退只带 usage 会把费用丢干净。
-        local out = { done = false }
-        if uses ~= nil then out.usage = uses end
-        if chunk.cost ~= nil then out.cost = tostring(chunk.cost) end
-        if uses == nil and chunk.cost == nil then return "" end
-        return json.encode(out)
-    end
-    local delta = chunk.choices[1].delta or {}
-    local fr = chunk.choices[1].finish_reason
-
-    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
-
-    local unified = {
-        content = delta.content or "",
-        done = (finish ~= nil)
-    }
-    if finish then
-        unified.finish_reason = finish
-    end
-    if uses ~= nil then
-        unified.usage = uses
-    end
-    -- zen 用 reasoning 字段承载推理文本（OpenAI 惯例是 reasoning_content）
-    local reasoning = delta.reasoning_content or delta.reasoning
-    if reasoning then
-        unified.reasoning_content = reasoning
-    end
-    if delta.tool_calls then
-        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
-        unified.tool_calls = delta.tool_calls
-    end
-    return json.encode(unified)
-end
-
--- 错误收敛（可选钩子）：zen 错误信封固定为 {error={type,message}}；
--- 免费池限流 FreeUsageLimitError 单独标注。返回 nil 走通用兜底。
-function adapter.transform_error(status, body)
-    local ok, resp = pcall(json.decode, body)
-    if not ok or type(resp) ~= "table" then return nil end
-    local e = resp.error
-    if type(e) ~= "table" then return nil end
-    if e.type == "FreeUsageLimitError" then
-        return "zen free pool quota exhausted"
-    end
-    return e.message
-end
-
-return adapter

internal/lua/adapters/qwen38.lua

@@ -0,0 +1,235 @@
+-- tokenplan / discovery-api (intern-ai.org.cn) — OpenAI-compatible with two
+-- differences that a stock OpenAI adapter gets wrong.
+--
+-- 1. REASONING IS A SEPARATE FIELD, AND IT IS MOST OF THE OUTPUT.
+--    Every response carries `reasoning_content` (non-stream) or
+--    `delta.reasoning_content` (stream), disjoint from `content`. Measured on
+--    qwen3.8-27b: a one-sentence answer cost completion=303 tokens of which
+--    reasoning=295 — 97% of the budget went to thinking. A caller handing the
+--    upstream's `content` straight through sees an empty string and concludes
+--    the model is broken, when in fact the answer was cut off by max_tokens
+--    before any of it became text.
+--
+--    So the reasoning is passed through as its own field rather than folded
+--    into content. Folding would be worse than dropping it: the page would
+--    render the model's private deliberation as its reply.
+--
+-- 2. `finish_reason: "length"` WITH `content: null` IS NORMAL HERE, NOT AN ERROR.
+--    A reasoning model that exhausts its budget before emitting text produces
+--    exactly `{"content": null, "reasoning_content": "...", "finish_reason":
+--    "length"}`. The generic empty-response guard in the gateway reads a
+--    content-less completion as "upstream returned nothing" and fails the
+--    attempt — which is why this adapter does not try to be clever: it reports
+--    both fields and the real finish_reason, and the caller decides.
+--
+-- 3. MODEL NAMES ARE REWRITTEN BY THE SERVER.
+--    Requesting `qwen3.8-27b` answers with `model: "qwen3.8-27b-fp8"`. The
+--    response's model field is the truth about who served the request, and the
+--    gateway records it (billing keys on the SERVED model, not the requested
+--    one). Nothing here normalises it back: rewriting it to the requested name
+--    would make the audit trail lie about what actually ran.
+--
+-- Also worth knowing when configuring this source: remote image URLs are
+-- fetched BY THE UPSTREAM and time out reliably (`media_fetch_failed`). Images
+-- must be sent inline as data: URLs. Measured: base64 worked, a Wikimedia URL
+-- did not.
+
+local adapter = {}
+
+adapter.name = "qwen38"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+local function clamp_reasoning(req)
+    -- Nothing to rewrite for reasoning: this endpoint has no `thinking` toggle
+    -- (unlike deepseek's extra_body.thinking). max_tokens is the only lever, and
+    -- raising it is the caller's decision, not the adapter's — silently
+    -- inflating a caller's token budget would change what they pay for.
+end
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok or type(req) ~= "table" then return raw_body end
+
+    -- Reasoning models here spend most of a small budget thinking and then get
+    -- truncated before any text. The failure is silent: the upstream returns
+    -- 200 with content=null, so it looks like a healthy request that produced
+    -- nothing. Warn rather than rewrite — see clamp_reasoning.
+    if type(req.max_tokens) == "number" and req.max_tokens > 0 and req.max_tokens < 512 then
+        clamp_reasoning(req)
+    end
+
+    -- An assistant message carrying tool_calls must echo reasoning_content back
+    -- on the next turn, or the upstream rejects the conversation. OpenAI-shaped
+    -- clients never send it, so an empty string is supplied the same way the
+    -- deepseek adapter does.
+    if type(req.messages) == "table" then
+        for _, msg in ipairs(req.messages) do
+            if type(msg) == "table" and msg.role == "assistant" then
+                if type(msg.tool_calls) == "table" and msg.tool_calls[1] ~= nil
+                    and msg.reasoning_content == nil then
+                    msg.reasoning_content = ""
+                end
+            end
+        end
+    end
+
+    return json.encode(req)
+end
+
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or type(resp) ~= "table" then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 },
+    }
+
+    if type(resp.usage) == "table" then
+        local u = resp.usage
+        unified.token_usage.prompt = u.prompt_tokens or 0
+        unified.token_usage.completion = u.completion_tokens or 0
+        unified.token_usage.total = u.total_tokens or 0
+        if type(u.prompt_tokens_details) == "table" and u.prompt_tokens_details.cached_tokens ~= nil then
+            unified.token_usage.prompt_tokens_details = { cached_tokens = u.prompt_tokens_details.cached_tokens }
+        end
+        -- reasoning_tokens is a SUBSET of completion_tokens, never additive.
+        --
+        -- It must go under completion_tokens_details, NOT at the top of
+        -- token_usage: that is the path internal/gateway/chat.go reads when it
+        -- fills the audit record's Reasoning field. Emitting it at the top level
+        -- type-checks fine and is then dropped at the one place that could have
+        -- kept it — which is how a reasoning model ends up billed as if it had
+        -- produced no reasoning at all.
+        --
+        -- (The deepseek adapter does not emit this field at all, so reasoning
+        -- billing has been reading zero everywhere so far. That is a separate
+        -- gap; this adapter does not depend on it being fixed.)
+        if type(u.completion_tokens_details) == "table"
+            and u.completion_tokens_details.reasoning_tokens ~= nil then
+            local d = unified.token_usage.completion_tokens_details
+            if type(d) ~= "table" then
+                d = {}
+                unified.token_usage.completion_tokens_details = d
+            end
+            d.reasoning_tokens = u.completion_tokens_details.reasoning_tokens
+        end
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            local m = ch.message
+            -- content may be nil on a truncated reasoning-only response; that is
+            -- reported as an empty string, NOT as a failure.
+            if type(m.content) == "string" then
+                unified.content = m.content
+            end
+            if type(m.reasoning_content) == "string" then
+                unified.reasoning_content = m.reasoning_content
+            end
+            if type(m.tool_calls) == "table" then
+                local tcs = {}
+                for _, tc in ipairs(m.tool_calls) do
+                    local fn = tc["function"] or {}
+                    local args_ok, args = pcall(json.decode, fn.arguments or "{}")
+                    if not args_ok then args = {} end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = fn.name,
+                        arguments = args,
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok or type(chunk) ~= "table" then return "" end
+
+    -- Usage arrives on a chunk with empty choices (the terminal usage chunk).
+    -- Keys must match Go's TokenUsage json tags; the gateway re-emits the
+    -- standard *_tokens fields.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        local u = chunk.usage
+        uses = {
+            prompt = u.prompt_tokens or 0,
+            completion = u.completion_tokens or 0,
+            total = u.total_tokens or 0,
+        }
+        if type(u.prompt_tokens_details) == "table" and u.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = u.prompt_tokens_details.cached_tokens }
+        end
+        -- Same nesting as the non-stream path, for the same reason: chat.go reads
+        -- reasoning under completion_tokens_details and nothing else.
+        if type(u.completion_tokens_details) == "table"
+            and u.completion_tokens_details.reasoning_tokens ~= nil then
+            uses.completion_tokens_details = {
+                reasoning_tokens = u.completion_tokens_details.reasoning_tokens
+            }
+        end
+    end
+
+    if type(chunk.choices) ~= "table" or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- A usage-only chunk is not a content or finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+
+    local ch = chunk.choices[1]
+    local delta = ch.delta or {}
+    local finish = (type(ch.finish_reason) == "string" and ch.finish_reason ~= "") and ch.finish_reason or nil
+
+    local unified = {
+        content = delta.content or "",
+        done = (finish ~= nil),
+    }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    -- The whole point of this adapter: reasoning is streamed as its own channel
+    -- and must stay separate. Measured: 20 reasoning chunks to 1 content chunk
+    -- on a one-sentence answer.
+    if type(delta.reasoning_content) == "string" then
+        unified.reasoning_content = delta.reasoning_content
+    end
+    if type(delta.tool_calls) == "table" then
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+-- Error envelope is standard OpenAI, but this endpoint also reports media
+-- failures with a distinct code (media_fetch_failed) when a remote image URL
+-- cannot be pulled. Both carry {error:{message,...}}; the message is surfaced
+-- with its code so "image fetch timed out" is not mistaken for a model problem.
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) ~= "table" then return nil end
+    if type(e.message) ~= "string" then return nil end
+    if type(e.code) == "string" and e.code ~= "" and e.code ~= "invalid_api_key" then
+        return e.message .. " (" .. e.code .. ")"
+    end
+    return e.message
+end
+
+return adapter

internal/lua/adapters/sensenova.lua

@@ -12,7 +12,13 @@ adapter.headers = {}
 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
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     if req.messages then
         for _, msg in ipairs(req.messages) do

internal/lua/adapters/trae.lua

@@ -15,7 +15,13 @@ adapter.headers = {}
 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
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     if req.messages then
         for _, msg in ipairs(req.messages) do

internal/lua/billing_test.go

@@ -2,12 +2,25 @@ package lua
 
 import (
 	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
 	"strings"
 	"testing"
 )
 
+// pluginLoaded reports whether a plugin with this name is present, ignoring any
+// LoadError: a plugin that is present but broken still counts as "loaded here",
+// and the caller decides whether the error matters.
+func pluginLoaded(ps *Plugins, name string) bool {
+	for _, p := range ps.List() {
+		if fmt.Sprint(p["name"]) == name {
+			return true
+		}
+	}
+	return false
+}
+
 // The billing plugin ships with the gateway, so its arithmetic is a contract:
 // a wrong price silently produces wrong money. These tests drive it through the
 // real hook path and check the NUMBERS, not merely that it loads.
@@ -447,8 +460,16 @@ return plugin`)
 	}
 
 	// Order must be honoured so the sidebar is predictable.
-	if len(ui.Pages) != 4 {
-		t.Fatalf("expected all 4 pages to merge, got %d: %v", len(ui.Pages), got)
+	//
+	// The page COUNT is not asserted: it used to be 4, and it grew every time a
+	// bundled plugin added a page. A hard count here reads as "the merge is
+	// broken" whenever the product gains a feature, which trains the reader to
+	// bump a number instead of reading the failure. What matters is that all
+	// three pages this test contributed merged, and that they came out sorted.
+	for _, want := range []string{"third-a", "third-b", "other-page"} {
+		if !got[want] {
+			t.Errorf("page %s did not merge: %v", want, got)
+		}
 	}
 	for i := 1; i < len(ui.Pages); i++ {
 		if ui.Pages[i].Order < ui.Pages[i-1].Order {
@@ -462,8 +483,15 @@ return plugin`)
 // configuration, since seeding only happens on a fresh plugin dir.
 func TestBillingPluginLoadedByDefault(t *testing.T) {
 	ps, pdir := billingVM(t)
-	if ps.Count() != 1 {
-		t.Fatalf("expected 1 bundled plugin, got %d", ps.Count())
+	// billing must be among the bundled plugins, not the only one. The count
+	// grows with every shipped plugin and asserting it makes this test a
+	// changelog detector.
+	if !pluginLoaded(ps, "billing") {
+		var names []string
+		for _, p := range ps.List() {
+			names = append(names, fmt.Sprint(p["name"]))
+		}
+		t.Fatalf("billing is not loaded; bundled plugins are %v", names)
 	}
 	if _, err := os.Stat(filepath.Join(pdir, "billing.lua")); err != nil {
 		t.Errorf("billing.lua was not written to the plugin dir: %v", err)
@@ -966,3 +994,245 @@ func TestBillingCacheHitClampedInStats(t *testing.T) {
 			tot["cache_fresh_tokens"])
 	}
 }
+
+// TestUnpricedIsTrackedPerDay pins the day-keyed unpriced tally.
+//
+// The lifetime counters cannot answer the question an operator actually has —
+// "is it STILL happening?" — because they keep the debt from before a rule was
+// added. After pricing a mispriced source the lifetime number never moves, so a
+// dashboard reading "unpriced: 6376" is indistinguishable from a system that is
+// still billing nothing. Day buckets are what make a pricing fix observable: the
+// offending day keeps its count and the next day starts at zero for it.
+func TestUnpricedIsTrackedPerDay(t *testing.T) {
+	ps, _ := billingVM(t)
+	// Nothing is priced in the default config, so every request is unpriced.
+	// Two requests on two different UTC days.
+	for _, ms := range []int64{1790000000000, 1790100000000} {
+		ps.Fire(StageRequestEnd, map[string]interface{}{
+			"model": "mystery-model", "source": "nowhere", "key": "***aaaaaa",
+			"ok": true, "prompt_tokens": 10, "completion_tokens": 5,
+			"time": ms,
+		})
+	}
+	st := stateOf(t, ps)
+	byDay, ok := st["unpriced_by_day"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("state has no unpriced_by_day map: %T", st["unpriced_by_day"])
+	}
+	if len(byDay) != 2 {
+		t.Fatalf("expected 2 day buckets, got %d: %v", len(byDay), byDay)
+	}
+	var sum float64
+	for day, bucket := range byDay {
+		b := bucket.(map[string]interface{})
+		sum += b["reqs"].(float64)
+		models := b["models"].(map[string]interface{})
+		if models["mystery-model"].(float64) != 1 {
+			t.Errorf("%s: model tally = %v, want 1", day, models["mystery-model"])
+		}
+	}
+	if sum != 2 {
+		t.Errorf("day buckets sum to %v, want 2 (the lifetime counter must not be double-counted)", sum)
+	}
+	// The lifetime counters still exist and still agree — this is additive, not
+	// a replacement.
+	if st["unpriced_reqs"].(float64) != 2 {
+		t.Errorf("unpriced_reqs = %v, want 2", st["unpriced_reqs"])
+	}
+}
+
+// TestReasoningBilledAtCompletionRateByDefault is the property that keeps the
+// standard formula matching a real invoice.
+//
+// reasoning_tokens is a subset of completion_tokens and every provider measured
+// (OpenAI, DeepSeek, Anthropic) charges it at the ordinary output rate. A rule
+// with no reasoning_price must therefore produce EXACTLY the number the
+// pre-reasoning formula produced — the field exists for a provider that prices
+// thinking differently, not as a new charge.
+func TestReasoningBilledAtCompletionRateByDefault(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				// USD PER TOKEN (compile.go divides the per-million figure by 1e6).
+				// prompt 0.15/M, completion 0.60/M. No reasoning_price anywhere.
+				"m": map[string]interface{}{"prompt": 0.15e-6, "completion": 0.60e-6},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000,
+		"reasoning_tokens": 800, // all of it thinking; still billed as output
+		"time":             1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	approx(t, "reasoning at completion rate", cost, 0.00075)
+}
+
+// TestReasoningPriceOverridesOnlyWhenSet: with a rule that DOES price thinking
+// differently, only the difference is applied. Charging the reasoning portion a
+// second time in full is the obvious wrong implementation and it doubles the
+// completion line.
+func TestReasoningPriceOverridesOnlyWhenSet(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 0.15e-6, "completion": 0.60e-6,
+					// Thinking costs 3x ordinary output: 1.80/M.
+					"reasoning_price": 1.80e-6,
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000, "reasoning_tokens": 800,
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	// base = 1000*0.15e-6 + 1000*0.60e-6 = 0.00075
+	//      + 800 * (1.80-0.60)e-6   = 0.00096  →  0.00171
+	approx(t, "reasoning repriced to 1.80/M", cost, 0.00171)
+}
+
+// TestReasoningClampedToCompletion: a provider claiming MORE reasoning than
+// completion is malformed. Without a clamp the visible-output count goes
+// negative and the delta exceeds the whole request, so the bill silently
+// explodes on bad data instead of degrading.
+func TestReasoningClampedToCompletion(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 0.15e-6, "completion": 0.60e-6, "reasoning_price": 1.80e-6,
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000, "reasoning_tokens": 99999,
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	// reasoning is clamped from 99999 down to completion (1000):
+	//   base  = 1000*0.15e-6 + 1000*0.60e-6            = 0.00075
+	//   delta = 1000 * (1.80-0.60)e-6                  = 0.00120
+	//   total                                          = 0.00195
+	// Without the clamp this would be 0.18 — 92x the whole request, and a
+	// number plausible enough to survive a glance at a dashboard.
+	approx(t, "reasoning clamped to completion", cost, 0.00195)
+}
+
+// TestLongContextStepIsMarginal: the threshold reprices only the tokens PAST
+// it, at a marginal rate.
+//
+// A flat prompt price under-bills exactly the traffic an agent workload
+// produces — on this deployment 88% of requests exceed 128k with a median of
+// 247k — and it under-bills silently, because the number stays plausible. The
+// step must apply to the excess only; charging the WHOLE prompt at the higher
+// rate is the other obvious implementation and it overcharges short requests by
+// the same factor it was meant to fix long ones.
+func TestLongContextStepIsMarginal(t *testing.T) {
+
+	// Short request: below the threshold, so the step must not touch it at all.
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 100000, "completion_tokens": 0, "time": 1750000000000,
+	})
+	got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
+	approx(t, "short request unaffected", got, 0.1)
+
+	// Long request: 200k prompt with a 128k threshold → 72k tokens at the step.
+	ps2, _ := billingVM(t)
+	if err := ps2.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps2.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 200000, "completion_tokens": 0, "time": 1750000000000,
+	})
+	got2 := stateOf(t, ps2)["total"].(map[string]interface{})["cost"].(float64)
+	// 128000*1e-6 + 72000*2e-6 = 0.128 + 0.144 = 0.272
+	approx(t, "long request marginal step", got2, 0.272)
+}
+
+// TestLongContextComposesWithCacheDiscount pins the interaction, which is the
+// whole reason for computing the step as a marginal rate.
+//
+// On this deployment the cache hit rate is 88% and long prompts dominate, so
+// "cached tokens beyond the threshold" is the common case, not an edge case. A
+// cache read already bills at prompt*discount; the long-context rate applies on
+// top of THAT, not on top of the full prompt price. Computing it any other way
+// silently changes the price of the majority of traffic.
+func TestLongContextComposesWithCacheDiscount(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD", "cache_discount": 0.1,
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 100000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	// 200k prompt, 150k of it cached, 50k fresh. Excess past 100k = 100k, split
+	// 25% fresh / 75% cached.
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 200000, "completion_tokens": 0,
+		"cache_hit_tokens": 150000, "cache_reported": true,
+		"time": 1750000000000,
+	})
+	got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
+	// base:  fresh 50000*1e-6            = 0.05
+	//        cache 150000*1e-6*0.1        = 0.015
+	// step:  fresh excess 25000*(2-1)e-6 = 0.025
+	//        cache excess 75000*(2-0.1)e-6 = 0.1425
+	// total                             = 0.2325
+	approx(t, "step composes with cache discount", got, 0.2325)
+}

internal/lua/billing_ui_test.go

@@ -414,3 +414,86 @@ func TestRuleTableLocksColumnWidthsInAColgroup(t *testing.T) {
 			"control and must never be squeezed by the flexible inputs")
 	}
 }
+
+// TestBillingUIShowsRecentUnpriced: the page must be able to distinguish a
+// settled pricing debt from a live leak. The lifetime unpriced counter answers
+// neither — it keeps every request from before a rule was added, so it reads the
+// same number in both situations. The day buckets exist to make the difference
+// visible, and a bucket nothing renders is a bucket nobody benefits from.
+func TestBillingUIShowsRecentUnpriced(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		"unpriced_by_day", // the tally is written
+		"recentUnpriced",  // it is reduced to a window
+		"unpriced_recent", // it reaches the view
+		"unpriced7d",      // the card labels it
+		"unpricedClean",   // and says so when the window is clear
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("billing UI lacks %q", needle)
+		}
+	}
+}
+
+// TestBillingRulesUIExposesReasoningPrice: the thinking-token rate must be
+// editable in the rules editor, not only settable by hand-editing config.yaml.
+// A pricing capability that lives only in the config file is one the operator
+// will assume does not exist.
+func TestBillingRulesUIExposesReasoningPrice(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		`class='m-r'`,         // the input, in both the existing rows and the "+ model" row
+		`models[n].reasoning`, // collected back out of the DOM on save
+		`reasoningHint`,       // and it says what leaving it blank means
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("rules editor lacks %q", needle)
+		}
+	}
+	// Two occurrences: an existing model's row and the template for a new one.
+	// One would mean the "+ model" button produces a row that silently drops
+	// the rate on the next save.
+	if n := strings.Count(src, `class='m-r'`); n < 2 {
+		t.Errorf("m-r appears %d time(s); both the row template and the editor need it", n)
+	}
+}
+
+// TestBillingRulesUIExposesLongContextStep: the long-context step must be
+// editable in the rules editor.
+//
+// This is not a niche control. On an agent workload 88% of requests carry more
+// than 128k prompt tokens with a median of 247k, so a rule that omits the step
+// under-bills the MAJORITY of traffic while every number on the dashboard stays
+// plausible. Leaving the capability in config.yaml alone would mean the operator
+// never discovers it needs it.
+func TestBillingRulesUIExposesLongContextStep(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		`class='m-lc-a'`,         // threshold
+		`class='m-lc-p'`,         // absolute rate
+		`class='m-lc-m'`,         // multiplier
+		`models[n].long_context`, // collected back out on save
+		`lcHint`,                 // explains that blank = no step
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("rules editor lacks %q", needle)
+		}
+	}
+	// Both the existing-row template and the "+ model" template need the three
+	// inputs; one would make a newly added model lose its step on the next save.
+	if n := strings.Count(src, `class='m-lc-a'`); n < 2 {
+		t.Errorf("long-context inputs appear %d time(s); both row templates need them", n)
+	}
+}

internal/lua/conflict_arbitration_test.go

@@ -0,0 +1,287 @@
+package lua
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// TestRouteConflictDetectedAndArbitrated（用户要求 A）：
+// 两个插件声明同一条 /api/plugins/http/ 服务路径 → 必须报冲突（列出双方），
+// 运维裁决后裁决生效；裁决持久化，重启（重新 LoadDir）后仍然生效。
+// 全程不崩溃、不拒绝加载 —— 两个插件都保持 loaded。
+func TestRouteConflictDetectedAndArbitrated(t *testing.T) {
+	dir := t.TempDir()
+	_, ps := bridgeVM(t)
+	ps.dir = dir
+
+	// a.lua 与 b.lua 都声明路径 "shared"
+	const aSrc = `
+local P = {}
+P.name = "plugA"
+P.version = "1"
+P.exports = {}
+P.http = { { path = "shared", handler = "h", roles = {"admin"} } }
+function P.h(req) return { status = 200, json = { from = "A" } } end
+return P
+`
+	const bSrc = `
+local P = {}
+P.name = "plugB"
+P.version = "1"
+P.exports = {}
+P.http = { { path = "shared", handler = "h", roles = {"admin"} } }
+function P.h(req) return { status = 200, json = { from = "B" } } end
+return P
+`
+	for name, src := range map[string]string{"plugA": aSrc, "plugB": bSrc} {
+		if err := ps.LoadSource(name, src); err != nil {
+			t.Fatalf("加载 %s: %v", name, err)
+		}
+	}
+
+	// ① 冲突被报出来，候选齐整，未裁决
+	cs := ps.RouteConflicts()
+	if len(cs) != 1 {
+		t.Fatalf("应恰有 1 条路径冲突, 实得 %d: %+v", len(cs), cs)
+	}
+	c := cs[0]
+	if c.Path != "shared" {
+		t.Fatalf("冲突路径应为 shared, 实得 %q", c.Path)
+	}
+	if len(c.Candidates) != 2 || c.Candidates[0] != "plugA" || c.Candidates[1] != "plugB" {
+		t.Fatalf("候选应为 [plugA plugB], 实得 %v", c.Candidates)
+	}
+	if c.Resolved != "" {
+		t.Fatalf("初始应未裁决, 实得 %q", c.Resolved)
+	}
+	// 两个插件都保持 loaded（不许因冲突拒绝启动/加载）
+	if ps.List()[0]["loaded"] != true || ps.List()[1]["loaded"] != true {
+		t.Fatal("冲突的两个插件都应保持 loaded")
+	}
+
+	// ② 非法裁决被拒（赢家必须确实是候选之一）
+	if err := ps.ResolveRouteConflict("shared", "plugC"); err == nil {
+		t.Fatal("赢家不是候选应被拒绝")
+	}
+	if err := ps.ResolveRouteConflict("other", "plugA"); err == nil {
+		t.Fatal("未冲突的路径不应可裁决")
+	}
+
+	// ③ 合法裁决 + 持久化
+	if err := ps.ResolveRouteConflict("shared", "plugB"); err != nil {
+		t.Fatalf("合法裁决失败: %v", err)
+	}
+	if ps.RouteWinner("shared") != "plugB" {
+		t.Fatalf("裁决后赢家应为 plugB, 实得 %q", ps.RouteWinner("shared"))
+	}
+	if _, err := os.Stat(filepath.Join(dir, ".route-conflicts.json")); err != nil {
+		t.Fatalf("裁决应落盘 .route-conflicts.json: %v", err)
+	}
+
+	// ④ 重启语义：新 registry 读回裁决
+	_, ps2 := bridgeVM(t)
+	ps2.dir = dir
+	if err := ps2.LoadRouteConflicts(); err != nil {
+		t.Fatalf("读回裁决: %v", err)
+	}
+	if err := ps2.LoadSource("plugA", aSrc); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps2.LoadSource("plugB", bSrc); err != nil {
+		t.Fatal(err)
+	}
+	cs2 := ps2.RouteConflicts()
+	if len(cs2) != 1 || cs2[0].Resolved != "plugB" {
+		t.Fatalf("重启后裁决应保留 plugB, 实得 %+v", cs2)
+	}
+}
+
+// TestRouteConflictWinnerGoneBackToUnresolved：已裁决的赢家插件被删/改后不再
+// 声明该路径 → 裁决作废回到未裁决态，而不是指向一个不存在的赢家。
+func TestRouteConflictWinnerGoneBackToUnresolved(t *testing.T) {
+	_, ps := bridgeVM(t)
+	const aSrc = `
+local P = {}
+P.name = "plugA"
+P.http = { { path = "s", handler = "h", roles = {"admin"} } }
+function P.h(req) return { status = 200, json = {} } end
+return P
+`
+	const bSrc = `
+local P = {}
+P.name = "plugB"
+P.http = { { path = "s", handler = "h", roles = {"admin"} } }
+function P.h(req) return { status = 200, json = {} } end
+return P
+`
+	if err := ps.LoadSource("plugA", aSrc); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadSource("plugB", bSrc); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.ResolveRouteConflict("s", "plugA"); err != nil {
+		t.Fatal(err)
+	}
+	// plugA 被禁用 → 不再占路由 → 冲突消失（不是回未裁决）
+	if err := ps.SetEnabled("plugA", false); err != nil {
+		t.Fatal(err)
+	}
+	if got := ps.RouteConflicts(); len(got) != 0 {
+		t.Fatalf("只 plugB 声明时不应再有冲突, 实得 %+v", got)
+	}
+}
+
+// TestAdapterNameConflictSuspendsForArbitration（用户要求 B）：
+// 插件安装与内置同名的适配器 → 不再硬拒绝，而是挂起待裁决：
+//   - 现任内置继续服务（版本号不变）
+//   - 返回 ErrAdapterConflictPending（typed，可区分「代码错」）
+//   - 冲突出现在 AdapterConflicts() 列表
+//   - 运维 ResolveAdapterConflict("incumbent") → 现任保留、冲突清除
+//   - 运维 ResolveAdapterConflict("candidate") → 插件版真正安装、归属转移
+//   - 裁决持久化；重启后插件再装同名适配器自动套用上次裁决
+func TestAdapterNameConflictSuspendsForArbitration(t *testing.T) {
+	dir := t.TempDir()
+	_, ps := bridgeVM(t)
+	ps.dir = dir
+
+	builtin := `local a={} a.name="openai" a.version="BUILTIN" a.endpoint="/x" return a`
+	if err := ps.vm.LoadAdapterSource("openai", builtin); err != nil {
+		t.Fatal(err)
+	}
+	evil := `local a={} a.name="openai" a.version="PLUGIN_V1" a.endpoint="/x" return a`
+
+	// ① 安装挂起：现任保留，typed 错误，不静默覆盖
+	got, err := ps.InstallAdapter("sneaky", "openai", evil)
+	if err == nil {
+		t.Fatal("撞内置名应返回错误（挂起）")
+	}
+	var pend *ErrAdapterConflictPending
+	if !errors.As(err, &pend) {
+		t.Fatalf("错误类型应为 ErrAdapterConflictPending, got %T: %v", err, err)
+	}
+	if got != "" {
+		t.Fatalf("挂起时不得返回生效名, 实得 %q", got)
+	}
+	if v := ps.vm.AdapterVersion("openai"); v != "BUILTIN" {
+		t.Fatalf("现任内置必须继续服务, 版本实得 %q", v)
+	}
+
+	// ② 冲突可见
+	ac := ps.AdapterConflicts()
+	if len(ac) != 1 || ac[0].Name != "openai" || ac[0].Candidate != "sneaky" {
+		t.Fatalf("挂起冲突应可见: %+v", ac)
+	}
+
+	// ③ 裁决 incumbent → 现任保留
+	if err := ps.ResolveAdapterConflict("openai", "incumbent"); err != nil {
+		t.Fatalf("裁决 incumbent: %v", err)
+	}
+	if v := ps.vm.AdapterVersion("openai"); v != "BUILTIN" {
+		t.Fatalf("裁决 incumbent 后内置应保留, 实得 %q", v)
+	}
+	if got := ps.AdapterConflicts(); len(got) != 0 {
+		t.Fatalf("裁决后冲突应清除, 实得 %+v", got)
+	}
+	// 归属仍是 operator（内置）
+	if o := ps.AdapterOwnerOf("openai"); o != "" {
+		t.Fatalf("内置归属应为空（operator）, 实得 %q", o)
+	}
+
+	// ④ 重启后插件再装 → 之前的 incumbent 裁决自动套用（不再二次挂起，也不安装）
+	if err := ps.LoadRouteConflicts(); err != nil {
+		t.Fatal(err)
+	}
+	_, err = ps.InstallAdapter("sneaky", "openai", evil)
+	var pend2 *ErrAdapterConflictPending
+	if !errors.As(err, &pend2) {
+		t.Fatalf("自动套用 incumbent 裁决应仍返回 pending 错误（未安装）, got %v", err)
+	}
+	if v := ps.vm.AdapterVersion("openai"); v != "BUILTIN" {
+		t.Fatalf("持久 incumbent 裁决应保持内置, 实得 %q", v)
+	}
+
+	// ⑤ 裁决 candidate → 插件版真正装上、归属转移、持久化
+	if err := ps.ResolveAdapterConflict("openai", "candidate"); err != nil {
+		t.Fatalf("裁决 candidate: %v", err)
+	}
+	if v := ps.vm.AdapterVersion("openai"); v != "PLUGIN_V1" {
+		t.Fatalf("裁决 candidate 后插件版应生效, 实得 %q", v)
+	}
+	if o := ps.AdapterOwnerOf("openai"); o != "sneaky" {
+		t.Fatalf("归属应转移给 sneaky, 实得 %q", o)
+	}
+	// 持久化文件里是 candidate:sneaky
+	b, _ := os.ReadFile(filepath.Join(dir, ".route-conflicts.json"))
+	if !strings.Contains(string(b), "candidate:sneaky") {
+		t.Fatalf("裁决应持久化为 candidate:sneaky, 文件: %s", b)
+	}
+}
+
+// TestAdapterPluginConflictSuspendsForArbitration：插件 A 装了适配器后插件 B
+// 再装同名 → 挂起待裁决（与内置撞名同一机制），现任 A 的版本继续服务。
+func TestAdapterPluginConflictSuspendsForArbitration(t *testing.T) {
+	_, ps := bridgeVM(t)
+	if _, err := ps.InstallAdapter("alpha", "shared_auth", tokenAdapterV1); err != nil {
+		t.Fatal(err)
+	}
+	_, err := ps.InstallAdapter("beta", "shared_auth", tokenAdapterV2)
+	var pend *ErrAdapterConflictPending
+	if !errors.As(err, &pend) {
+		t.Fatalf("B 装同名应挂起, got %v", err)
+	}
+	// tokenAdapterV1 声明 name="myauth"，池落在生效名下（注册键 shared_auth）。
+	if v := ps.vm.AdapterVersion("myauth"); v != "1.0" {
+		t.Fatalf("alpha 的现任版本应继续服务, 实得 %q", v)
+	}
+	ac := ps.AdapterConflicts()
+	if len(ac) != 1 || ac[0].Incumbent != "plugin:alpha" {
+		t.Fatalf("冲突应点名现任归属 plugin:alpha, 实得 %+v", ac)
+	}
+	// 裁决 candidate → beta 接管
+	if err := ps.ResolveAdapterConflict("shared_auth", "candidate"); err != nil {
+		t.Fatal(err)
+	}
+	if v := ps.vm.AdapterVersion("myauth"); v != "2.0" {
+		t.Fatalf("beta 版应生效, 实得 %q", v)
+	}
+	if o := ps.AdapterOwnerOf("myauth"); o != "beta" {
+		t.Fatalf("归属应转移给 beta, 实得 %q", o)
+	}
+}
+
+// TestAdapterConflictRejectsBadWinner：裁决值必须合法；无挂起冲突时拒绝。
+func TestAdapterConflictRejectsBadWinner(t *testing.T) {
+	_, ps := bridgeVM(t)
+	if err := ps.ResolveAdapterConflict("nope", "candidate"); err == nil {
+		t.Fatal("无挂起冲突应拒绝")
+	}
+	if _, err := ps.InstallAdapter("p", "myauth", tokenAdapterV1); err != nil {
+		t.Fatal(err)
+	}
+	// myauth 是插件自己装的 → 更新路径（update 走 UpdateAdapter，非冲突）
+	if err := ps.ResolveAdapterConflict("myauth", "bogus"); err == nil {
+		t.Fatal("非法 winner 值应拒绝")
+	}
+}
+
+// TestPluginOwnUpdateStillWorks：插件更新自己装的适配器（token 轮换）不走冲突
+// 路径 —— 行为与改造前一致。守住「挂起只发生在撞名」这条边界。
+func TestPluginOwnUpdateStillWorks(t *testing.T) {
+	_, ps := bridgeVM(t)
+	if _, err := ps.InstallAdapter("authtoken", "myauth", tokenAdapterV1); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.UpdateAdapter("authtoken", "myauth", tokenAdapterV2); err != nil {
+		t.Fatalf("own update 不应受影响: %v", err)
+	}
+	if v := ps.vm.AdapterVersion("myauth"); v != "2.0" {
+		t.Fatalf("轮换后版本应 2.0, 实得 %q", v)
+	}
+	if got := ps.AdapterConflicts(); len(got) != 0 {
+		t.Fatalf("own update 不应产生冲突, 实得 %+v", got)
+	}
+}

internal/lua/kernel_api.go

@@ -0,0 +1,498 @@
+package lua
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+)
+
+// Package lua — 插件可调用的内核动作面（R4-c）
+//
+// 用户需求（原话）：「现在这个接口不全，可以多给一些接口让 lua 调用内核。」
+//
+// 在此之前，插件只能通过 __llmsproxy_adapters 管适配器 —— 那是三个手写闭包。
+// 内核另有十几个公开动作（重载配置、清冷却、重探上游、增删源/密钥、改 AUTO 链…），
+// 插件一个都调不到。而"自愈 / 联动 / 自动孵化"这类插件能力恰恰需要它们：
+//
+//	探测到上游整片挂掉 → probe_all + reset_health
+//	发现额度变化       → save_auto_rules 重排链路
+//	拿到新端点         → add_source
+//
+// # 为什么 Go 侧 deadline 是可行的（而 Lua 死循环不可中断）
+//
+// 这是整个设计的支点。两者形态完全不同：
+//
+//	Lua 里 while true do end            → 无法中断（LuaJIT 限制，实测三次）
+//	Lua 里 __llmsproxy_kernel.call(...)  → **可以**超时
+//
+// 因为后者 Lua 只是在等一个返回值，不占用 Lua 栈的执行权。所以每次调用在 Go
+// 侧包一层 select + ctx deadline：超时就把 (nil, err) 还给 Lua。这比"Lua 投递
+// 请求、Go 队列执行"干净 —— 队列还要定义"超时后积压的命令算什么"，而 deadline
+// 的语义是闭合的。
+//
+// # 权限：全开，但默认不授予
+//
+// 用户选定「全开」（13 个动作都可达）+ 「设置页可配权限」。全开描述的是**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。
+// 因此默认拒绝，授权记录在 .access.json 里（复用现有覆盖机制）。
+//
+// danger 级动作必须**逐个**授权，不接受通配符 —— 否则运维图省事写个 "*"，
+// "删密钥"就和"读状态"同权了。
+
+// KernelAPI 是内核暴露给插件的能力面。由 internal/core 实现并注入。
+//
+// 依赖方向：internal/lua 不能 import internal/core（方向是 core → lua），
+// 所以这里定义接口、core 侧实现 —— 与 SetOutboundHosts 同一个注入先例。
+type KernelAPI interface {
+	// Actions 列出可调用的动作。设置页据此渲染权限开关，插件据此发现能力。
+	Actions() []KernelAction
+	// Invoke 执行一个动作。实现**必须**尊重 ctx：ctx 取消后应尽快返回，
+	// 否则插件侧的超时虽然会返回错误，但该动作仍在后台跑（变成不可见的泄漏）。
+	Invoke(ctx context.Context, name string, args map[string]interface{}) (interface{}, error)
+}
+
+// 危险级。它决定权限授予的严格程度，也决定设置页用什么颜色提示。
+const (
+	// DangerRead 无副作用，只读。仍然需要授权，但可以一次授予全部。
+	DangerRead = "read"
+	// DangerWrite 修改运行状态或配置，但可恢复（重载、清冷却、改链路）。
+	DangerWrite = "write"
+	// DangerDestructive 删除类，或让服务中断。**必须逐个授权**。
+	DangerDestructive = "destructive"
+)
+
+// KernelAction 描述一个可调用动作。
+type KernelAction struct {
+	// Deferred marks an action that MUST NOT run while the Lua caller holds its
+	// plugin lock. The Lua entry points (hook, exported call, scheduled task) all
+	// hold p.mu for the whole call, and an action that re-enters the plugin layer
+	// (set_plugin_enabled reaches SetEnabled → p.mu.Lock()) would self-deadlock —
+	// and with a deadline, would then execute invisibly AFTER the plugin was told
+	// it failed.
+	//
+	// Such an action is queued and run on its own goroutine once the Lua call has
+	// returned, so the caller gets "accepted" rather than a lock-up or a lie.
+	Deferred bool
+
+	Name        string   `json:"name"`
+	Title       string   `json:"title"`
+	Description string   `json:"description"`
+	Danger      string   `json:"danger"`
+	Params      []string `json:"params,omitempty"`
+	// TimeoutMS is the Go-side deadline for one invocation. Long actions (a
+	// full re-probe walks every source over the network) need a bigger budget
+	// than a state read; making it per-action rather than one global number is
+	// what keeps the deadline honest rather than arbitrary.
+	TimeoutMS int `json:"timeout_ms"`
+}
+
+// kernelCatalog is the full action set. It lives in the lua package (rather than
+// being discovered from the implementation) so that permissions can be validated
+// even when no Core is wired — tests, and a gateway with plugins but no kernel.
+//
+// Every action here maps to an existing Core method; adding one is a deliberate
+// act rather than a reflex. ListKeys/FindKey are deliberately ABSENT: they return
+// PLAINTEXT gateway keys, and handing those to third-party Lua would hand over
+// the deployment's entire credential surface. Key-related reads go through
+// masked forms instead.
+var kernelCatalog = []KernelAction{
+	// ---- 只读 ----
+	{Name: "status", Title: "运行状态", Danger: DangerRead, TimeoutMS: 3000,
+		Description: "各源的调度状态（冷却、连续失败、最近错误）"},
+	{Name: "list_sources", Title: "源列表", Danger: DangerRead, TimeoutMS: 3000,
+		Description: "已配置的上游源（api_key 已掩码）"},
+	{Name: "list_models", Title: "模型列表", Danger: DangerRead, TimeoutMS: 3000,
+		Description: "当前可路由的模型 id 列表"},
+	{Name: "list_adapters", Title: "适配器列表", Danger: DangerRead, TimeoutMS: 3000,
+		Description: "已加载的适配器名与版本"},
+	{Name: "list_auto_rules", Title: "AUTO 链路", Danger: DangerRead, TimeoutMS: 3000,
+		Description: "当前全局 AUTO 链的槽位"},
+	{Name: "adapter_owner", Title: "适配器归属", Danger: DangerRead, TimeoutMS: 3000,
+		Params:      []string{"name"},
+		Description: "某适配器由谁安装（operator 或 plugin:名字）"},
+
+	// ---- 写（可恢复） ----
+	{Name: "reload", Title: "重载配置", Danger: DangerWrite, TimeoutMS: 30000,
+		Description: "重新读取运行配置并重建源/链路"},
+	{Name: "reset_health", Title: "清除冷却", Danger: DangerWrite, TimeoutMS: 5000,
+		Description: "清掉所有源的退避状态（修好上游后立即恢复）"},
+	{Name: "probe_all", Title: "全量探测", Danger: DangerWrite, TimeoutMS: 60000,
+		Description: "对每个源做一次真实可达性探测（走网络，慢）"},
+	{Name: "save_auto_rules", Title: "设置 AUTO 链", Danger: DangerWrite, TimeoutMS: 10000,
+		Params:      []string{"entries"},
+		Description: "替换全局 AUTO 链；entries 为 [{model, source, ...}] 数组"},
+	{Name: "save_key_auto", Title: "设置按 key 链路", Danger: DangerWrite, TimeoutMS: 10000,
+		Params:      []string{"key_name", "entries"},
+		Description: "替换某个网关 key 自己的 AUTO 链"},
+	{Name: "add_source", Title: "新增源", Danger: DangerWrite, TimeoutMS: 30000,
+		Params:      []string{"source"},
+		Description: "把一个新的上游源接入（source 为源定义表）"},
+	{Name: "create_key", Title: "新建网关密钥", Danger: DangerWrite, TimeoutMS: 10000,
+		Params:      []string{"name", "role", "models"},
+		Description: "创建一个网关 key（返回的明文密钥只在此次响应里出现）"},
+
+	// ---- 破坏性：必须逐个授权 ----
+	{Name: "remove_source", Title: "删除源", Danger: DangerDestructive, TimeoutMS: 30000,
+		Params:      []string{"name"},
+		Description: "删除一个上游源"},
+	{Name: "delete_key", Title: "删除网关密钥", Danger: DangerDestructive, TimeoutMS: 10000,
+		Params:      []string{"key"},
+		Description: "删除一个网关 key（正在使用的客户端会立刻 401）"},
+	{Name: "update_key", Title: "修改网关密钥", Danger: DangerDestructive, TimeoutMS: 10000,
+		Params:      []string{"key", "name", "role", "models"},
+		Description: "改动 key 的角色/可见模型（可能扩大访问面）"},
+	{Name: "remove_adapter", Title: "删除适配器", Danger: DangerDestructive, TimeoutMS: 30000,
+		Params:      []string{"name"},
+		Description: "删除适配器（仅限插件自己装的；内置与非自有的会被拒）"},
+	{Name: "set_plugin_enabled", Title: "启停插件", Danger: DangerDestructive, TimeoutMS: 5000,
+		Params: []string{"name", "enabled"}, Deferred: true,
+		Description: "启用或停用某个插件（异步执行：在本次调用返回后生效）"},
+}
+
+// kernelActionByName indexes the catalog.
+var kernelActionByName = func() map[string]KernelAction {
+	m := make(map[string]KernelAction, len(kernelCatalog))
+	for _, a := range kernelCatalog {
+		m[a.Name] = a
+	}
+	return m
+}()
+
+// KernelActions returns the catalog sorted by name (stable for the UI).
+func KernelActions() []KernelAction {
+	out := append([]KernelAction(nil), kernelCatalog...)
+	sort.Slice(out, func(i, j int) bool { return out[i].Name < out[j].Name })
+	return out
+}
+
+// LookupKernelAction finds an action, reporting whether it exists.
+func LookupKernelAction(name string) (KernelAction, bool) {
+	a, ok := kernelActionByName[name]
+	return a, ok
+}
+
+// ---- 权限 ----
+
+// errNotPermitted is returned when a plugin calls an action it was not granted.
+// It names the action and the plugin so the operator can act on it directly.
+func notPermitted(plugin, action string) error {
+	return fmt.Errorf("kernel action %q is not granted to plugin %q; "+
+		"enable it in the plugin's kernel permissions", action, plugin)
+}
+
+// KernelPermission is the operator's grant for one plugin: the set of action
+// names it may call, plus whether it was granted "everything readable".
+type KernelPermission struct {
+	// Actions is the explicit allowlist, already expanded (no wildcards stored,
+	// so a NEW action added in a future release is NOT silently granted to an
+	// existing plugin — that is the safe direction for a security default).
+	Actions []string
+	// AllRead is true when the operator granted every read-only action at once.
+	// Stored separately from Actions so that a future read-only action IS covered
+	// without re-saving, while writes still require an explicit grant.
+	AllRead bool
+}
+
+// clone returns a copy so callers cannot mutate the stored permission.
+func (p KernelPermission) clone() KernelPermission {
+	return KernelPermission{Actions: append([]string(nil), p.Actions...), AllRead: p.AllRead}
+}
+
+// allows reports whether the permission covers action.
+//
+// The "all readable" grant deliberately does NOT cover a future WRITE action:
+// expanding a wildcard into whatever happens to exist later is how a convenience
+// grant turns into a privilege escalation.
+func (p KernelPermission) allows(a KernelAction) bool {
+	if p.AllRead && a.Danger == DangerRead {
+		return true
+	}
+	for _, n := range p.Actions {
+		if n == a.Name {
+			return true
+		}
+	}
+	return false
+}
+
+// ---- 调用（带 Go 侧超时） ----
+
+// ErrKernelUnavailable is returned when no KernelAPI was injected (a gateway
+// running plugins without a wired core, e.g. some tests).
+var ErrKernelUnavailable = errors.New("kernel actions are not available in this build")
+
+// callKernel invokes one action with the permission check and the Go-side
+// deadline, returning a value Lua can consume.
+//
+// The deadline is the whole point: a plugin can block here harmlessly, because
+// Lua is waiting on a return value rather than spinning. On timeout the caller
+// gets a normal error; the action's own goroutine is told to stop via ctx.
+func (ps *Plugins) callKernel(plugin, name string, args map[string]interface{}) (interface{}, error) {
+	api := ps.kernel()
+	if api == nil {
+		return nil, ErrKernelUnavailable
+	}
+	action, ok := LookupKernelAction(name)
+	if !ok {
+		return nil, fmt.Errorf("unknown kernel action %q", name)
+	}
+	if !ps.kernelPermitted(plugin, action) {
+		return nil, notPermitted(plugin, name)
+	}
+
+	timeout := time.Duration(action.TimeoutMS) * time.Millisecond
+	if timeout <= 0 {
+		timeout = 10 * time.Second
+	}
+	// The caller's identity is injected HERE, not taken from args: an action that
+	// needs to know who is calling (remove_adapter, to enforce ownership) must not
+	// be able to be lied to by the plugin passing someone else's name.
+	if args == nil {
+		args = map[string]interface{}{}
+	}
+	args["plugin"] = plugin
+
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+
+	type result struct {
+		v   interface{}
+		err error
+	}
+	// A deferred action must not run under the caller's plugin lock, so it is
+	// queued and returns "accepted" immediately. Reporting success here without
+	// waiting is honest: the caller is told the action was ACCEPTED, and the
+	// actual outcome lands in the plugin's hook_errors surface if it fails.
+	if action.Deferred {
+		ps.runDeferred(action, args)
+		return map[string]interface{}{
+			"accepted": true,
+			"note":     "this action runs asynchronously; check the plugin state/errors for the outcome",
+		}, nil
+	}
+
+	// The action runs on its own goroutine so this call can give up on time.
+	// Buffered so a late-finishing action never blocks on sending.
+	done := make(chan result, 1)
+	go func() {
+		defer func() {
+			// A panicking kernel action must not take the gateway down through a
+			// Lua call. Convert it into an error the plugin sees.
+			if rec := recover(); rec != nil {
+				done <- result{err: fmt.Errorf("kernel action %s panicked: %v", name, rec)}
+			}
+		}()
+		v, err := api.Invoke(ctx, name, args)
+		done <- result{v: v, err: err}
+	}()
+
+	select {
+	case r := <-done:
+		return r.v, r.err
+	case <-ctx.Done():
+		return nil, fmt.Errorf("kernel action %q exceeded its %s budget", name, timeout)
+	}
+}
+
+// runDeferred executes a lock-re-entrant action outside the Lua call.
+//
+// It runs on a fresh goroutine precisely so it can take the plugin lock that the
+// caller still holds; waiting for the caller to unwind first would need a
+// handshake that is more machinery than the one action needs. A failure is
+// surfaced through hook_errors, which is where an operator already looks for
+// "a plugin's background work stopped working".
+func (ps *Plugins) runDeferred(a KernelAction, args map[string]interface{}) {
+	api := ps.kernel()
+	if api == nil {
+		return
+	}
+	timeout := time.Duration(a.TimeoutMS) * time.Millisecond
+	if timeout <= 0 {
+		timeout = 10 * time.Second
+	}
+	// Copy the args: the caller's map is still owned by the Lua call frame.
+	cp := make(map[string]interface{}, len(args))
+	for k, v := range args {
+		cp[k] = v
+	}
+	go func() {
+		defer func() {
+			if rec := recover(); rec != nil {
+				ps.hookErr.note(StageRequestEnd, fmt.Sprintf("deferred kernel action %s panicked: %v", a.Name, rec))
+			}
+		}()
+		ctx, cancel := context.WithTimeout(context.Background(), timeout)
+		defer cancel()
+		if _, err := api.Invoke(ctx, a.Name, cp); err != nil {
+			ps.hookErr.note(StageRequestEnd, fmt.Sprintf("deferred kernel action %s failed: %v", a.Name, err))
+		}
+	}()
+}
+
+// kernelPermitted applies the stored grant, with one deliberate exception: a
+// plugin's OWN adapter may always be removed by that plugin (it is not a kernel
+// resource, it is one the plugin created). Everything else needs a grant.
+func (ps *Plugins) kernelPermitted(plugin string, a KernelAction) bool {
+	if a.Name == "remove_adapter" {
+		return true
+	}
+	ps.mu.RLock()
+	perm := ps.kernelPerms[plugin]
+	ps.mu.RUnlock()
+	return perm.allows(a)
+}
+
+// SetPluginKernelPermission replaces one plugin's grant and persists it.
+//
+// Validation happens here rather than at call time so a typo'd action name is
+// reported to the operator immediately instead of silently granting nothing.
+func (ps *Plugins) SetPluginKernelPermission(plugin string, actions []string, allRead bool) error {
+	if ps.find(plugin) == nil {
+		return fmt.Errorf("plugin %s not found", plugin)
+	}
+	clean := make([]string, 0, len(actions))
+	seen := map[string]bool{}
+	for _, raw := range actions {
+		name := strings.TrimSpace(raw)
+		if name == "" || seen[name] {
+			continue
+		}
+		a, ok := LookupKernelAction(name)
+		if !ok {
+			return fmt.Errorf("unknown kernel action %q", name)
+		}
+		seen[name] = true
+		clean = append(clean, a.Name)
+	}
+	sort.Strings(clean)
+	ps.mu.Lock()
+	if ps.kernelPerms == nil {
+		ps.kernelPerms = kernelPermissionSet{}
+	}
+	if len(clean) == 0 && !allRead {
+		delete(ps.kernelPerms, plugin)
+	} else {
+		ps.kernelPerms[plugin] = KernelPermission{Actions: clean, AllRead: allRead}
+	}
+	ps.saveKernelPermissionsLocked()
+	ps.mu.Unlock()
+	return nil
+}
+
+// PluginKernelPermission returns a copy of one plugin's grant.
+func (ps *Plugins) PluginKernelPermission(plugin string) KernelPermission {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	return ps.kernelPerms[plugin].clone()
+}
+
+// KernelPermissionView is the UI-facing form: what the plugin ASKED for (from its
+// own source) next to what the operator GRANTED. The distinction matters — a
+// plugin declaring `plugin.kernel = {...}` is a request, not a privilege.
+type KernelPermissionView struct {
+	Plugin   string   `json:"plugin"`
+	Declared []string `json:"declared,omitempty"`
+	Granted  []string `json:"granted,omitempty"`
+	AllRead  bool     `json:"all_read"`
+}
+
+// KernelPermissions lists every plugin's declared-vs-granted state for the UI.
+func (ps *Plugins) KernelPermissions() []KernelPermissionView {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	out := make([]KernelPermissionView, 0, len(ps.plugins))
+	for _, p := range ps.plugins {
+		perm := ps.kernelPerms[p.Info.Name]
+		out = append(out, KernelPermissionView{
+			Plugin:   p.Info.Name,
+			Declared: append([]string(nil), p.KernelDeclared...),
+			Granted:  append([]string(nil), perm.Actions...),
+			AllRead:  perm.AllRead,
+		})
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Plugin < out[j].Plugin })
+	return out
+}
+
+// kernel returns the injected API, or nil.
+func (ps *Plugins) kernel() KernelAPI {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	return ps.kernelAPI
+}
+
+// SetKernelAPI injects the kernel action implementation. Called during wiring,
+// mirroring SetOutboundHosts.
+func (ps *Plugins) SetKernelAPI(api KernelAPI) {
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	ps.kernelAPI = api
+}
+
+// kernelPermissionSet is the persisted per-plugin grant map.
+type kernelPermissionSet map[string]KernelPermission
+
+// jsonPermission is the on-disk form of one grant. Kept separate from
+// KernelPermission so the persisted shape cannot drift with a struct refactor
+// (the access file is read by older and newer binaries across an upgrade).
+type jsonPermission struct {
+	Actions []string `json:"actions,omitempty"`
+	AllRead bool     `json:"all_read,omitempty"`
+}
+
+func (p KernelPermission) marshal() jsonPermission {
+	return jsonPermission{Actions: p.Actions, AllRead: p.AllRead}
+}
+
+func (j jsonPermission) unmarshal() KernelPermission {
+	return KernelPermission{Actions: j.Actions, AllRead: j.AllRead}
+}
+
+type jsonPluginPolicy struct {
+	Actions []string `json:"actions,omitempty"`
+	AllRead bool     `json:"all_read,omitempty"`
+	Hosts   []string `json:"hosts,omitempty"`
+}
+
+func policyFromActions(p KernelPermission) jsonPluginPolicy {
+	return jsonPluginPolicy{Actions: p.Actions, AllRead: p.AllRead}
+}
+
+// KernelActionView is one action as the settings page needs it: the catalog entry
+// plus whether THIS plugin currently holds it and whether the plugin asked for it.
+type KernelActionView struct {
+	KernelAction
+	Granted  bool `json:"granted"`
+	Declared bool `json:"declared"`
+}
+
+// KernelActionsView lists the catalog annotated for one plugin.
+//
+// Granted and Declared are both reported because they answer different questions:
+// "may this plugin do it" and "did this plugin ask for it". An action that is
+// declared but not granted is the one an operator most needs to notice.
+func (ps *Plugins) KernelActionsView(plugin string) []KernelActionView {
+	perm := ps.PluginKernelPermission(plugin)
+	declared := map[string]bool{}
+	ps.mu.RLock()
+	for _, p := range ps.plugins {
+		if p.Info.Name == plugin {
+			for _, d := range p.KernelDeclared {
+				declared[d] = true
+			}
+		}
+	}
+	ps.mu.RUnlock()
+	out := make([]KernelActionView, 0, len(kernelCatalog))
+	for _, a := range KernelActions() {
+		out = append(out, KernelActionView{
+			KernelAction: a,
+			Granted:      perm.allows(a),
+			Declared:     declared[a.Name],
+		})
+	}
+	return out
+}

internal/lua/kernel_perms_persist.go

@@ -0,0 +1,126 @@
+package lua
+
+import (
+	"encoding/json"
+	"os"
+)
+
+// Kernel action grants are OPERATOR policy, persisted separately from both
+// plugin state and the endpoint/visibility overrides.
+//
+// Why a separate file rather than reusing .access.json:
+//
+//   - .access.json is a flat map[string]string ("name" or "name#path" -> spec).
+//     Kernel grants are a list plus a boolean, so folding them in would either
+//     change the shape of a file an older binary still reads, or force an
+//     encoding that cannot round-trip. An upgrade must not corrupt a running
+//     deployment's restrictions.
+//   - The two answer different questions — "who may call this endpoint" vs
+//     "what may this plugin do to the kernel" — and an operator looking at one
+//     should not have to parse the other.
+//
+// Shape: { "<plugin>": { "actions": ["reload", ...], "all_read": true } }
+
+const kernelPermsFileName = ".kernel-perms.json"
+
+func (ps *Plugins) kernelPermsFile() string {
+	if ps.dir == "" {
+		return ""
+	}
+	return ps.dir + string(os.PathSeparator) + kernelPermsFileName
+}
+
+// LoadKernelPermissions reads the persisted grants. A missing file is the normal
+// fresh-install state (nothing granted); a corrupt file is logged and ignored
+// rather than fatal — losing a grant is recoverable, refusing to boot is not.
+func (ps *Plugins) LoadKernelPermissions() error {
+	path := ps.kernelPermsFile()
+	if path == "" {
+		return nil
+	}
+	b, err := os.ReadFile(path)
+	if os.IsNotExist(err) {
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	// The file holds BOTH kernel-action grants (map[string]jsonPermission, the
+	// historical shape) AND, for plugins that asked for them, the operator's
+	// extra-host grants (map[string]jsonPluginPolicy.Hosts). A bare
+	// {actions:[...]} row still parses into jsonPluginPolicy with empty Hosts,
+	// so old files load unchanged.
+	raw := map[string]jsonPluginPolicy{}
+	if err := json.Unmarshal(b, &raw); err != nil {
+		return err
+	}
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	if ps.kernelPerms == nil {
+		ps.kernelPerms = kernelPermissionSet{}
+	}
+	if ps.extraHosts == nil {
+		ps.extraHosts = map[string][]string{}
+	}
+	for name, jp := range raw {
+		ps.kernelPerms[name] = KernelPermission{Actions: jp.Actions, AllRead: jp.AllRead}
+		if len(jp.Hosts) > 0 {
+			// Load the persisted grant into extraHosts WITHOUT folding it into the
+			// allowlist yet — sources are loaded later, and the caller owns that
+			// fold. GrantExtraHosts rebuilds the allowlist when sources are ready.
+			ps.extraHosts[name] = append([]string(nil), jp.Hosts...)
+		}
+	}
+	return nil
+}
+
+// saveKernelPermissionsLocked flushes the grant map. Caller holds ps.mu.
+//
+// Grants drop unknown action names on write. If an operator authorized an action
+// that a later release removed, keeping the name would leave a stale entry that
+// nothing can ever satisfy — and it would reappear in the UI as if it were live.
+//
+// Host grants travel in the same file (jsonPluginPolicy.Hosts): a granted host is
+// the operator's explicit decision to let a plugin egress somewhere, and losing
+// it across a restart (the previous behaviour) made grants silently vanish.
+func (ps *Plugins) saveKernelPermissionsLocked() {
+	path := ps.kernelPermsFile()
+	if path == "" {
+		return
+	}
+	out := map[string]jsonPluginPolicy{}
+	something := false
+	for name, perm := range ps.kernelPerms {
+		valid := make([]string, 0, len(perm.Actions))
+		for _, a := range perm.Actions {
+			if _, ok := LookupKernelAction(a); ok {
+				valid = append(valid, a)
+			}
+		}
+		row := policyFromActions(KernelPermission{Actions: valid, AllRead: perm.AllRead})
+		if hs := ps.extraHosts[name]; len(hs) > 0 {
+			row.Hosts = append([]string(nil), hs...)
+		}
+		if len(valid) == 0 && !perm.AllRead && len(row.Hosts) == 0 {
+			continue // nothing granted: presence would be a lie
+		}
+		out[name] = row
+		something = true
+	}
+	// A plugin may hold host grants with NO kernel actions granted (operators
+	// approve hosts and actions independently). Emit those too.
+	for name, hs := range ps.extraHosts {
+		if len(hs) == 0 {
+			continue
+		}
+		if _, exists := out[name]; !exists {
+			out[name] = jsonPluginPolicy{Hosts: append([]string(nil), hs...)}
+			something = true
+		}
+	}
+	if !something {
+		_ = os.Remove(path)
+		return
+	}
+	writeJSONAtomic(path, out)
+}

internal/lua/persist_test.go

@@ -281,3 +281,60 @@ func TestCloseFlushesTail(t *testing.T) {
 		t.Errorf("flushed n = %v, want 1", saved.State["n"])
 	}
 }
+
+// TestOnDemandCallIsPersisted guards the persistence gap that the quota memo fell
+// into, and it needs a plugin with NO hooks to see it.
+//
+// markDirty fired from two places: hooks and SetState. `Call` — the on-demand path
+// a plugin page uses — fired from neither, so a plugin that does all its work in
+// exported functions (quota: it declares no hooks at all) mutated state that
+// nothing ever wrote. The visible symptom was a health probe that looked inert
+// from outside: the memo changed in memory and the file kept the previous
+// generation until some unrelated configuration write happened to flush it.
+//
+// A hookless plugin is the whole point of the fixture: the counter above
+// accumulates through a hook, which is the path that already persisted, so it
+// cannot see this defect.
+func TestOnDemandCallIsPersisted(t *testing.T) {
+	ps, _ := newPersistVM(t)
+	if err := ps.LoadSource("ondemand", `
+local plugin = { name = "ondemand", version = "1.0" }
+plugin.state = { n = 0 }
+plugin.exports = { "bump" }
+function plugin.bump(payload)
+  plugin.state.n = plugin.state.n + 1
+  return { n = plugin.state.n }
+end
+return plugin
+`); err != nil {
+		t.Fatal(err)
+	}
+
+	for i := 1; i <= 3; i++ {
+		out, err := ps.Call("ondemand", "bump", map[string]interface{}{})
+		if err != nil {
+			t.Fatalf("call %d: %v", i, err)
+		}
+		if m, _ := out.(map[string]interface{}); m["n"] != float64(i) {
+			t.Fatalf("call %d returned %v, want n=%d", i, out, i)
+		}
+	}
+	// Close() waits for the final flush, so no timing is involved: anything the
+	// calls did NOT mark dirty cannot appear here.
+	ps.Close()
+
+	b, err := os.ReadFile(ps.stateFile("ondemand"))
+	if err != nil {
+		t.Fatalf("★ an exported call did not mark the plugin dirty: %v — a plugin "+
+			"with no hooks can never persist its work", err)
+	}
+	var saved struct {
+		State map[string]interface{} `json:"state"`
+	}
+	if err := json.Unmarshal(b, &saved); err != nil {
+		t.Fatal(err)
+	}
+	if saved.State["n"] != float64(3) {
+		t.Errorf("persisted n = %v after three exported calls, want 3", saved.State["n"])
+	}
+}

internal/lua/plugin_access.go

@@ -0,0 +1,180 @@
+package lua
+
+import "strings"
+
+// One vocabulary for "who may see / read / call this".
+//
+// Before this there were two rules with OPPOSITE defaults, living side by side:
+//
+//   - audienceAllows(want, role):  "" or "all" → everybody
+//     Used by UI visibility and per-page/per-element audiences.
+//   - HTTPHandler.allows(role, key): "" → admin ONLY
+//     Used by registered endpoints, where "I forgot to declare" must not mean
+//     "open to the internet".
+//
+// Both defaults are right for their own surface, and having them in one file
+// with one name is how a reader ends up assuming the other one's rule applies.
+// An operator configuring "who can see billing" and "who can call /balance" had
+// to learn two vocabularies to answer the same question.
+//
+// The distinction that remains is not the RULE, it is the DEFAULT, so it is
+// stated as data rather than left implicit: an empty grant is "deny" for
+// capabilities and "inherit" for visibility, and every call site names which it
+// wants.
+
+// AccessRule is a parsed grant: a set of roles, an optional set of named gateway
+// keys, and an explicit public flag.
+//
+// Keys REPLACE roles rather than adding to them. That is not an arbitrary
+// choice: "allow these two keys" read as "these two keys AND everyone in the
+// user role" silently widens access to every user key, which is the opposite of
+// what someone tightening a rule meant. `keys = { "@alice" }` therefore means
+// exactly alice.
+type AccessRule struct {
+	// Roles admit a gateway role: "admin", "user", …
+	Roles []string
+	// Keys admit a named gateway key, by CONFIG NAME (never the secret itself, so
+	// a rule file contains no credential and survives key rotation).
+	Keys []string
+	// Public admits anyone, including an unauthenticated caller.
+	Public bool
+}
+
+// AccessAll is the grant that admits everyone.
+func AccessAll() AccessRule { return AccessRule{Public: true} }
+
+// ParseAccessRule reads the one grant syntax the plugin API accepts:
+//
+//	"admin"                        roles
+//	"admin, user"                  roles
+//	"@alice, @bob"                 named keys, replacing the role rule
+//	"user, @alice"                 the named keys plus the roles (union)
+//	"public"                       anyone
+//	"all"                          anyone (the older spelling of public)
+//	""                             empty — meaning depends on the caller
+//
+// An "@name" token is a key, anything else is a role. That is unambiguous to
+// type and to parse, and it means one string can express everything the rules
+// editor needs to offer.
+//
+// "all" is kept as an alias for "public" because it is what existing plugin
+// files and saved operator overrides already contain. Dropping it would have
+// silently turned a published grant into a closed one.
+func ParseAccessRule(s string) AccessRule {
+	var r AccessRule
+	for _, part := range strings.Split(s, ",") {
+		part = strings.TrimSpace(part)
+		if part == "" {
+			continue
+		}
+		if strings.EqualFold(part, "public") || strings.EqualFold(part, "all") {
+			r.Public = true
+			continue
+		}
+		if strings.HasPrefix(part, "@") {
+			if k := strings.TrimSpace(strings.TrimPrefix(part, "@")); k != "" {
+				r.Keys = append(r.Keys, k)
+			}
+			continue
+		}
+		r.Roles = append(r.Roles, part)
+	}
+	return r
+}
+
+// Rule renders the grant back to the syntax ParseAccessRule accepts. The admin
+// page uses it to fill its input, so a value that round-trips is a value the
+// operator can edit without translating.
+func (r AccessRule) Rule() string {
+	var parts []string
+	for _, k := range r.Keys {
+		parts = append(parts, "@"+k)
+	}
+	for _, role := range r.Roles {
+		parts = append(parts, role)
+	}
+	if r.Public {
+		parts = append(parts, "public")
+	}
+	return strings.Join(parts, ", ")
+}
+
+// Empty reports whether the grant names nobody.
+func (r AccessRule) Empty() bool {
+	return len(r.Roles) == 0 && len(r.Keys) == 0 && !r.Public
+}
+
+// AllowsCapability reports whether a CAPABILITY (an endpoint, a callable
+// function) is reachable. Empty means DENY: a plugin that forgets to declare
+// gets a 403 naming the field to set, never an open door.
+//
+// Admin is always admitted. An operator has to be able to inspect what a plugin
+// exposes even while locked out of it, or a typo in a key list would leave
+// nobody able to diagnose the endpoint they were trying to tighten.
+func (r AccessRule) AllowsCapability(role, keyName string) bool {
+	if role == "admin" {
+		return true
+	}
+	if len(r.Keys) > 0 {
+		// A non-empty key list REPLACES the role rule (see the type comment).
+		// An unauthenticated caller has no key name, so a key-scoped capability
+		// is unreachable for them — even when the same rule also lists roles,
+		// because a caller with no key is also not the kind of caller a role
+		// rule was written for.
+		if keyName == "" {
+			return false
+		}
+		for _, k := range r.Keys {
+			if strings.EqualFold(k, keyName) {
+				return true
+			}
+		}
+		// The key list replaced the role rule, so a miss is a deny even though
+		// roles may also be listed.
+		return false
+	}
+	if r.Public {
+		return true
+	}
+	for _, want := range r.Roles {
+		if strings.EqualFold(want, role) {
+			return true
+		}
+	}
+	return false
+}
+
+// AllowsVisibility reports whether a VIEW (a page, a UI element) is rendered.
+// Empty means INHERIT: visibility is additive by nature — a page nobody can see
+// is simply not rendered — so an undeclared audience must not hide a page the
+// plugin clearly intends to publish.
+func (r AccessRule) AllowsVisibility(role string) bool {
+	if r.Empty() || r.Public {
+		return true
+	}
+	if role == "admin" && len(r.Keys) > 0 {
+		// A key-scoped view still has to be inspectable by an admin, for the
+		// same reason capabilities do.
+		return true
+	}
+	for _, want := range r.Roles {
+		if strings.EqualFold(want, role) {
+			return true
+		}
+	}
+	return false
+}
+
+// accessRuleFor folds the separate declaration fields of an endpoint into one
+// grant, so the rules below live in one place instead of beside each field.
+//
+// It is a plain fold, deliberately. An earlier version also reclassified a role
+// that happened to start with "@" into a key, on the theory that "roles =
+// { '@alice' }" is a natural mistake. That is a behaviour change smuggled into a
+// refactor: it would have quietly widened who reaches an endpoint whose author
+// wrote something nonsensical. If that mistake is worth fixing, it is a change
+// with its own justification and its own test, not a side effect of unifying
+// two functions into one.
+func accessRuleFor(roles, keys []string, public bool) AccessRule {
+	return AccessRule{Roles: roles, Keys: keys, Public: public}
+}

internal/lua/plugin_access_persist.go

@@ -0,0 +1,115 @@
+package lua
+
+import (
+	"encoding/json"
+	"os"
+	"sort"
+)
+
+// Access overrides are OPERATOR policy, not plugin data. They live in their own
+// file rather than in each plugin's state because:
+//
+//   - they describe the deployment, not the plugin, so a plugin upgrade that
+//     replaces its .state.json must not silently widen access again;
+//   - they must survive a restart. An in-memory map made "restrict billing to
+//     admin" look like it worked right up to the next deploy, at which point the
+//     restriction evaporated with nothing in the logs to say why;
+//   - the plugin_dir is the one place an operator already looks when asking
+//     "what does this plugin actually get to see", so the answer lives beside it.
+//
+// The file is a flat map of "name" (whole plugin) or "name#path" (one endpoint)
+// to an access spec, which is the same shape the in-memory store uses, so load
+// and save cannot disagree about what a key means.
+
+const accessFileName = ".access.json"
+
+func (ps *Plugins) accessFile() string {
+	if ps.dir == "" {
+		return ""
+	}
+	return ps.dir + string(os.PathSeparator) + accessFileName
+}
+
+// LoadAccessOverrides reads the persisted overrides into memory. A missing file
+// is not an error: no operator has restricted anything yet, which is the
+// default state of a fresh install.
+//
+// A corrupt file is also not fatal. Losing a restriction is bad, but refusing to
+// start the gateway is worse, and an empty map plus a logged error leaves the
+// operator with something to fix.
+func (ps *Plugins) LoadAccessOverrides() error {
+	path := ps.accessFile()
+	if path == "" {
+		return nil
+	}
+	b, err := os.ReadFile(path)
+	if os.IsNotExist(err) {
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	m := map[string]string{}
+	if err := json.Unmarshal(b, &m); err != nil {
+		return err
+	}
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	if ps.visibility == nil {
+		ps.visibility = visibilityOverrides{}
+	}
+	for k, v := range m {
+		ps.visibility[k] = v
+	}
+	return nil
+}
+
+// saveAccessOverridesLocked flushes the whole override map. It is called on
+// every change, and unlike plugin state the map is tiny and changes only when an
+// operator clicks Save — so a synchronous write costs nothing that matters, and
+// it means the change is durable the moment the API returns 200.
+func (ps *Plugins) saveAccessOverridesLocked() {
+	path := ps.accessFile()
+	if path == "" {
+		return
+	}
+	out := make(map[string]string, len(ps.visibility))
+	for k, v := range ps.visibility {
+		if v != "" {
+			out[k] = v
+		}
+	}
+	if len(out) == 0 {
+		// Leave no empty file behind: its absence is the honest encoding of
+		// "nothing is overridden".
+		_ = os.Remove(path)
+		return
+	}
+	writeJSONAtomic(path, out)
+}
+
+// AccessOverrides returns a copy of every override in force, for the UI.
+func (ps *Plugins) accessOverridesSnapshot() map[string]string {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	if len(ps.visibility) == 0 {
+		return nil
+	}
+	out := make(map[string]string, len(ps.visibility))
+	for k, v := range ps.visibility {
+		out[k] = v
+	}
+	return out
+}
+
+// sortedOverrideKeys exists only so a debug dump is stable; map iteration order
+// would otherwise make two identical states look different.
+func (ps *Plugins) sortedOverrideKeys() []string {
+	m := ps.accessOverridesSnapshot()
+	keys := make([]string, 0, len(m))
+	for k := range m {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	return keys
+}