test(retry): 钉住声明式重试判据与流式握手重试语义

- config：ShouldRetry/DelayDuration 表驱动（nil/次数=1/状态不匹配/body 不匹配/
  空子串/多信号/非法间隔）
- provider：命中策略就重试直到 200、body 不匹配**不许**重试（400 也常用于真错误）、
  流式只重试握手（首帧之后不重试，避免重复输出）

Makefile

@@ -1,4 +1,4 @@
-.PHONY: all build test gui gui-dev gui-dist gui-deb gui-win gui-win-docker help clean
+.PHONY: all build test gui gui-dev gui-dist gui-deb gui-win gui-win-docker help clean gui-test
 
 # ModelRouter - single-binary gateway + optional Electron desktop GUI (embedded core)
 export BINARY
@@ -65,6 +65,13 @@ gui-dist-dir:
 test:
 	go test ./...
 
+# Desktop shell unit tests (no Electron needed): plugin-tree seeding/migration.
+# These guard the two things that silently break the desktop build:
+#   - plugin_dir missing from an embedded profile  -> channels plugin invisible
+#   - bundled plugin files never refreshed         -> old channels.lua survives upgrades
+gui-test:
+	@cd $(GUI_DIR) && node --test plugin-sync.test.js
+
 clean:
 	rm -rf $(BUILD_DIR) cmd/build cmd/gui/dist cmd/gui/build/gui-dist cmd/gui/bin
 
@@ -75,6 +82,7 @@ help:
 	@echo "  make gui        run desktop GUI (dev)"
 	@echo "  make gui-dev    run GUI with devtools"
 	@echo "  make gui-dist   build deb + rpm + AppImage (linux)"
+	@echo "  make gui-test   desktop shell unit tests (plugin tree seeding)"
 	@echo "  make gui-deb    build deb only"
 	@echo "  make gui-win    build Windows nsis (needs host mingw + wine)"
 	@echo "  make gui-win-docker  build Windows nsis all-in-docker (no host mingw needed)"

cmd/gui/main.js

@@ -18,6 +18,7 @@ const http = require("http");
 const os = require("os");
 const crypto = require("crypto");
 const { spawn, execFileSync } = require("child_process");
+const pluginSync = require("./plugin-sync");
 
 const LOG_FILE = path.join(app.getPath("userData"), "gui.log");
 function log(msg) {
@@ -40,6 +41,13 @@ console.error = function () {
 // Per-user profile dir: <userData>/profile
 const PROFILE_DIR = path.join(app.getPath("userData"), "profile");
 const CONFIG_FILE = path.join(PROFILE_DIR, "config.yaml");
+// Plugin tree for the embedded core. The core deliberately does NOT default
+// plugin_dir (an operator opts in), so the desktop build must set it — without
+// it the whole plugin system stays off and the channels plugin (multi-platform
+// account login + credential vault + auto check-in) is invisible in the WebUI.
+// Seeding is the core's job (SeedBundled fills MISSING files only), so a user's
+// edited copy always wins.
+const PLUGINS_DIR = path.join(PROFILE_DIR, "plugins");
 // Packaged: extraResource at process.resourcesPath/bin (outside asar, spawnable).
 // Dev:   cmd/gui/bin, populated by prepare-core.js.
 const CORE_EXE = path.join(
@@ -216,6 +224,9 @@ function writeDefaultEmbeddedConfig() {
     "listen: 127.0.0.1:" + (settings.port || DEFAULT_PORT),
     "default_model: AUTO",
     "adapter_dir: " + JSON.stringify(path.join(PROFILE_DIR, "adapters")),
+    // Plugin tree: channels/billing/quota/usage ship inside the core binary and
+    // are seeded into this dir on first start. Omitting it disables plugins.
+    "plugin_dir: " + JSON.stringify(PLUGINS_DIR),
     "runtime_file: " + JSON.stringify(path.join(PROFILE_DIR, "runtime.json")),
     "",
     // Admin key as a plain `keys` entry (no seed flag) so the web UI never
@@ -268,6 +279,9 @@ function startCore() {
   // ensure profile dir + config + local adapters exist
   fs.mkdirSync(PROFILE_DIR, { recursive: true });
   if (!fs.existsSync(CONFIG_FILE)) writeDefaultEmbeddedConfig();
+  pluginSync.ensurePluginDirConfigured(CONFIG_FILE, PLUGINS_DIR, log);
+  fs.mkdirSync(PLUGINS_DIR, { recursive: true });
+  pluginSync.syncBundledPluginsOnUpgrade(PROFILE_DIR, PLUGINS_DIR, app.getVersion(), log);
 
   coreProc = spawn(CORE_EXE, ["-config", CONFIG_FILE], {
     cwd: PROFILE_DIR,
@@ -622,10 +636,55 @@ function createWindow() {
     },
   });
   mainWindow.loadFile(path.join(__dirname, "renderer", "index.html"));
+  // External links ALWAYS go to the user's default browser, never an in-app
+  // window. Three paths have to be covered, and missing any one of them is why
+  // "click the login link and nothing happens / the app navigates away" shows up:
+  //   1. window.open / target=_blank            -> setWindowOpenHandler
+  //   2. top-level navigation (same-tab click)  -> will-navigate
+  //   3. navigation inside the WebUI <iframe>   -> will-frame-navigate
+  // The embedded core's own origin (and file:// for the shell) must keep
+  // working, otherwise the WebUI and plugin pages stop loading.
+  const isInternal = (url) =>
+    !url ||
+    url === "about:blank" ||
+    url.startsWith("file://") ||
+    url.startsWith(embeddedBaseUrl()) ||
+    url.startsWith("devtools://");
+  const openExternally = (url) => {
+    if (url && (url.startsWith("http://") || url.startsWith("https://"))) {
+      log("[gui] open in default browser: " + url);
+      shell.openExternal(url);
+    }
+  };
   mainWindow.webContents.setWindowOpenHandler(({ url }) => {
-    if (url && url.startsWith("http")) shell.openExternal(url);
+    if (!isInternal(url)) openExternally(url);
     return { action: "deny" };
   });
+  mainWindow.webContents.on("will-navigate", (e, url) => {
+    if (isInternal(url)) return;
+    e.preventDefault();
+    openExternally(url);
+  });
+  // iframe navigations (the WebUI content area and the plugin pages inside it).
+  // The channels plugin renders its server-side login browser (noVNC) in an
+  // iframe on a LOOPBACK port, so loopback MUST stay in-app — blocking it would
+  // break VNC login. Everything non-loopback is external: open the default
+  // browser instead of silently replacing the gateway UI with a third-party page.
+  const isLoopback = (url) => {
+    try {
+      const h = new URL(url).hostname;
+      return h === "127.0.0.1" || h === "localhost" || h === "[::1]";
+    } catch (e) {
+      return false;
+    }
+  };
+  if (mainWindow.webContents.on) {
+    mainWindow.webContents.on("will-frame-navigate", (e) => {
+      if (isInternal(e.url) || isLoopback(e.url)) return;
+      e.preventDefault();
+      openExternally(e.url);
+    });
+  }
   // Silent start: `--silent` CLI flag or settings.silentStart keeps the app
   // in the tray without showing the window (Clash-Verge style). First run
   // always shows the window so the user sees the dashboard.

cmd/gui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.3",
+  "version": "2.0.46",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "modelrouter-gui",
-      "version": "1.9.3",
+      "version": "2.0.46",
       "license": "Proprietary",
       "devDependencies": {
         "electron": "^33.0.0",

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.7",
+  "version": "2.0.46",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",
@@ -81,6 +81,7 @@
     },
     "files": [
       "main.js",
+      "plugin-sync.js",
       "preload.js",
       "renderer/**/*",
       "build/**/*",

cmd/gui/plugin-sync.js

@@ -0,0 +1,82 @@
+// Plugin-tree maintenance for the embedded core (no Electron imports, so this
+// is unit-testable with plain node).
+//
+// Two problems this solves, both invisible from the gateway side:
+//   1. The core deliberately does NOT default `plugin_dir` (an operator opts
+//      in). A desktop profile without it runs with ZERO plugins — the channels
+//      plugin (account login, credential vault, auto check-in) never appears.
+//   2. The core seeds bundled plugins only when the file is ABSENT, so a
+//      desktop upgrade would keep the user's old channels.lua forever and
+//      silently skip every channel fix. On version change the shell must drop
+//      the built-in files so the new binary reseeds them.
+const fs = require("fs");
+const path = require("path");
+
+const BUILTIN_PLUGINS = ["billing.lua", "channels.lua", "quota.lua", "usage.lua"];
+
+// ensurePluginDirConfigured back-fills `plugin_dir` into an existing config.
+// Appending a top-level YAML key is safe for the shapes this file can have
+// (flat mapping + a keys list). Returns "added" | "present" | "absent" | "error".
+function ensurePluginDirConfigured(configFile, pluginsDir, log) {
+  try {
+    if (!fs.existsSync(configFile)) return "absent";
+    const cur = fs.readFileSync(configFile, "utf-8");
+    if (/^\s*plugin_dir\s*:/m.test(cur)) return "present";
+    fs.writeFileSync(
+      configFile,
+      cur.replace(/\s*$/, "") +
+        "\n# added by ModelRouter Desktop: enables the plugin tree (channels etc.)\n" +
+        "plugin_dir: " + JSON.stringify(pluginsDir) + "\n",
+      "utf-8",
+    );
+    if (log) log("[gui] plugin_dir back-filled into " + configFile);
+    return "added";
+  } catch (e) {
+    if (log) log("[gui] ensurePluginDirConfigured failed: " + e.message);
+    return "error";
+  }
+}
+
+// syncBundledPluginsOnUpgrade drops built-in plugin files when the app version
+// changed, so the core reseeds them from the binary it shipped with. Returns
+// { changed, dropped:[...], kept: "prev-<ver>" }.
+function syncBundledPluginsOnUpgrade(profileDir, pluginsDir, version, log) {
+  const stamp = path.join(profileDir, ".plugin-seed-version");
+  let have = "";
+  try {
+    have = fs.readFileSync(stamp, "utf-8").trim();
+  } catch (e) {
+    /* first run */
+  }
+  if (have === version) return { changed: false, dropped: [] };
+  const dropped = [];
+  for (const name of BUILTIN_PLUGINS) {
+    const f = path.join(pluginsDir, name);
+    if (!fs.existsSync(f)) continue;
+    try {
+      if (have) fs.renameSync(f, f + ".prev-" + have);
+      else fs.unlinkSync(f);
+      dropped.push(name);
+    } catch (e) {
+      if (log) log("[gui] drop builtin plugin " + name + " failed: " + e.message);
+    }
+  }
+  try {
+    fs.writeFileSync(stamp, version + "\n", "utf-8");
+  } catch (e) {
+    if (log) log("[gui] write plugin seed stamp failed: " + e.message);
+  }
+  if (dropped.length && log) {
+    log(
+      "[gui] plugins refreshed for " + version + " (" + dropped.length +
+        " builtin file(s) reseeded by core)",
+    );
+  }
+  return { changed: true, dropped, kept: have ? "prev-" + have : "" };
+}
+
+module.exports = {
+  BUILTIN_PLUGINS,
+  ensurePluginDirConfigured,
+  syncBundledPluginsOnUpgrade,
+};

cmd/gui/plugin-sync.test.js

@@ -0,0 +1,80 @@
+// Unit tests for the desktop plugin-tree maintenance (plain node, no Electron).
+// Run: node --test cmd/gui/plugin-sync.test.js
+const test = require("node:test");
+const assert = require("node:assert");
+const fs = require("fs");
+const os = require("os");
+const path = require("path");
+const { BUILTIN_PLUGINS, ensurePluginDirConfigured, syncBundledPluginsOnUpgrade } = require("./plugin-sync");
+
+const tmp = () => fs.mkdtempSync(path.join(os.tmpdir(), "plugin-sync-"));
+
+test("back-fills plugin_dir into an existing profile without it", () => {
+  const dir = tmp();
+  const cfg = path.join(dir, "config.yaml");
+  fs.writeFileSync(cfg, 'listen: 127.0.0.1:8787\ndefault_model: AUTO\nkeys:\n  - key: k\n    role: admin\n');
+  const plugins = path.join(dir, "plugins");
+  assert.equal(ensurePluginDirConfigured(cfg, plugins, null), "added");
+  const out = fs.readFileSync(cfg, "utf-8");
+  assert.match(out, /^plugin_dir: "/m);
+  assert.ok(out.includes(plugins.replace(/\\/g, "\\\\")) || out.includes(plugins), "path written");
+  // the rest of the profile must survive untouched
+  assert.match(out, /^listen: 127\.0\.0\.1:8787$/m);
+  assert.match(out, /^  - key: k$/m);
+});
+
+test("leaves a profile that already sets plugin_dir alone", () => {
+  const dir = tmp();
+  const cfg = path.join(dir, "config.yaml");
+  const orig = 'plugin_dir: "/custom/plugins"\nlisten: ":1"\n';
+  fs.writeFileSync(cfg, orig);
+  assert.equal(ensurePluginDirConfigured(cfg, "/whatever", null), "present");
+  assert.equal(fs.readFileSync(cfg, "utf-8"), orig);
+});
+
+test("missing config is not created by the back-fill", () => {
+  const dir = tmp();
+  const cfg = path.join(dir, "config.yaml");
+  assert.equal(ensurePluginDirConfigured(cfg, path.join(dir, "p"), null), "absent");
+  assert.equal(fs.existsSync(cfg), false);
+});
+
+test("upgrade drops builtin plugins (keeping a backup) but not user plugins", () => {
+  const dir = tmp();
+  const plugins = path.join(dir, "plugins");
+  fs.mkdirSync(plugins, { recursive: true });
+  for (const f of BUILTIN_PLUGINS) fs.writeFileSync(path.join(plugins, f), "OLD");
+  fs.writeFileSync(path.join(plugins, "my-own.lua"), "MINE");
+  fs.writeFileSync(path.join(dir, ".plugin-seed-version"), "1.0.0\n");
+
+  const r = syncBundledPluginsOnUpgrade(dir, plugins, "2.0.42", null);
+  assert.equal(r.changed, true);
+  assert.deepEqual(r.dropped.sort(), [...BUILTIN_PLUGINS].sort());
+  for (const f of BUILTIN_PLUGINS) {
+    assert.equal(fs.existsSync(path.join(plugins, f)), false, f + " dropped");
+    assert.equal(fs.readFileSync(path.join(plugins, f + ".prev-1.0.0"), "utf-8"), "OLD");
+  }
+  assert.equal(fs.readFileSync(path.join(plugins, "my-own.lua"), "utf-8"), "MINE");
+  assert.equal(fs.readFileSync(path.join(dir, ".plugin-seed-version"), "utf-8").trim(), "2.0.42");
+});
+
+test("same version is a no-op (nothing dropped)", () => {
+  const dir = tmp();
+  const plugins = path.join(dir, "plugins");
+  fs.mkdirSync(plugins, { recursive: true });
+  fs.writeFileSync(path.join(plugins, "channels.lua"), "CURRENT");
+  fs.writeFileSync(path.join(dir, ".plugin-seed-version"), "2.0.42\n");
+  const r = syncBundledPluginsOnUpgrade(dir, plugins, "2.0.42", null);
+  assert.equal(r.changed, false);
+  assert.equal(fs.readFileSync(path.join(plugins, "channels.lua"), "utf-8"), "CURRENT");
+});
+
+test("first run on a fresh profile just stamps the version", () => {
+  const dir = tmp();
+  const plugins = path.join(dir, "plugins");
+  fs.mkdirSync(plugins, { recursive: true });
+  const r = syncBundledPluginsOnUpgrade(dir, plugins, "2.0.42", null);
+  assert.equal(r.changed, true);
+  assert.deepEqual(r.dropped, []);
+  assert.equal(fs.existsSync(path.join(dir, ".plugin-seed-version")), true);
+});

docs/channel-issues-2026-10-19.md

@@ -12,7 +12,7 @@
   `TRAE_SOLO_ADAPTER_SRC`（channels.lua:9744），两份代码不同步。）
 - **修法**：`TRAE_SOLO_ADAPTER_SRC.transform_response` 加 SSE 聚合（复用流式帧解析，
   聚合成一个非流式响应），并把 `internal/lua/adapters/trae.lua` 的聚合逻辑对齐。
-- **状态**：⬜
+- **状态**：✅ 已修复（`aggregate_solo_sse`）并部署生产，实测非流式返回 200 正常内容。
 
 ## P2 buddy（CodeBuddy）非流式 400（11101 Non-stream not supported）
 - **现象**：`stream:false` 打 `channels-buddy:*` 上游 400 `11101: Non-stream chat request is currently not supported`。
@@ -20,7 +20,7 @@
   非流式请求原样透传被上游拒绝。buddy 的 `Accept: text/event-stream` 头已发，但 body.stream 仍是 false。
 - **修法**：`buddy.lua transform_request` 强制 `req.stream = true`；`transform_response` 增加
   SSE 聚合兜底（内核非流式路径收到 SSE 时聚合）。参考 opencodezen.lua 的既有做法（门禁第 4 条）。
-- **状态**：⬜
+- **状态**：✅ 已修复并部署生产，实测非流式返回 200 正常内容。
 
 ## P3 buddy 偶发 provider busy（429）＋ 11133 参数拒绝
 - **现象**：连续请求 `channels-buddy` 间歇返回 `provider busy`（source 并发槽 max_concurrent=1
@@ -43,7 +43,7 @@
   空窗期 DOM 仍挂旧渠道账号行；且慢的旧请求返回会覆盖新渠道数据。
 - **修法**：① renderDetail 立即清空账号/模型区并置"加载中"；② renderAccounts/renderModels
   加代际校验（`state.selected !== id` 时丢弃响应）。
-- **状态**：✅ 已改（channels.lua），待部署。
+- **状态**：✅ 已修复并部署生产。
 
 ## P6 CodeArts「完全不可用」——多数模型实际可用
 - **现象**：用户报"codearts 完全不可用"。实测逐模型探测：
@@ -55,29 +55,68 @@
   基础设施丢了挂接点）。需要把 on_request_end 补上 dead-model 记录。
 - **凭据**：expiresAt 已过期（15:05），`account/refresh` 续期成功（refreshable=true）。
   建议把「凭据到期自动续期」纳入定时任务检查。
-- **状态**：⬜（待补 dead-model 挂接 + GLM-5.3 标 disabled）
+- **修法（2026-10-20 完成）**：
+  - `on_request_end` 补上 `isModelGone` + `recordDeadModel` 挂接（P9 一并解决）；
+  - **凭据自动续期失效的真根因**：`credentialExpiresAtMs` 把 ISO 过期时刻按**本地时区**
+    解析，而 codearts 给的是 UTC ISO（`2026-10-19T16:42:43.06Z`）⇒ 本机 +08 下过期时刻
+    虚早 8 小时 ⇒ `refresh_creds` 每 30 分钟都判"已过期"并对华为云真打一次 STS
+    （token 本来还有 23 小时寿命）。已改为按声明时区解析（Z/±hh:mm/无后缀=UTC 三种形态
+    同一 epoch），回归测试 `TestCredentialExpiresAtMsParsesISOAsUTC`。
+- **状态**：✅ 已修复并部署生产。
 
 ## P7 CodeBuddy 海外账号 / Qoder 国内账号
 - **现象**：CodeBuddy 登录只走国内（copilot.tencent.com），Qoder 登录只走国际
   （openapi.qoder.sh / qoder.com）。用户要求两边都能选区。
 - **根因**：平台表单区硬编码。buddy 的 state-poll 端点与 Qoder 的 PKCE 端点都锁死一个区域。
-- **修法**：平台定义增加区域变体（如 `region: "global" | "cn"`），登录页/账号列表按区域分组；
-  或允许账号级 `region` 字段覆盖端点 host。
-- **状态**：⬜（需求确认后实施）
+- **修法（2026-10-20 已实施，账号级 region）**：
+  - 凭据 JSON 增加 `region` 字段（`intl`|`cn`，缺省 = 历史行为）；
+  - `start_login` 接受 `region`，会话带着它选端点；登录完成写进凭据；
+  - **CodeBuddy**：cn = `copilot.tencent.com`（platform=ide），intl = `www.workbuddy.ai`
+    （platform=`workbuddy-ai`，X-Domain/X-Product-Code/UA 按区域切换；实测
+    `POST /v2/plugin/auth/state?platform=workbuddy-ai` 回 200 信封）；
+  - **Qoder**：intl = `openapi.qoder.sh` + `qoder.com`（client_id e883ade2-…），
+    cn = `openapi.qoder.com.cn` + `qoder.cn`（client_id **不同**：732aef47-…，
+    照参考 QODER_CN；用国际版 id 授权页报「参数无效」）。CN 无公开 chat 端点
+    （参考实测 503）⇒ CN 账号只做登录/领取，不产源；
+  - UI：buddy/qoder 详情页多一个「＋国际区」按钮（默认国内/国际主区）。
+- **状态**：✅ 已实施并部署生产。
 
-## P8 商汤（Raccoon）扫码登录「链接无效」
-- **现象**：用户点登录后看到的二维码/链接打不开。
-- **排查**：登录会话本身正常发起（二维码已生成、轮询端点存在）。
-  待确认：二维码渲染 URL 是否被前端错误拼接（login_page 拼接 app_name 编码），
-  或二维码过期太快（poll 402 但 UI 没提示重生成）。
-- **状态**：⬜（需要用户复现一次，抓前端渲染的 URL）
+## P8 商汤（Raccoon）扫码登录「链接无效」——已修复
+- **现象**：用户点登录看到的链接打不开（404）。
+- **根因（两个，2026-10-20 实测定位）**：
+  ① 登录页 `/login/mp` **已被站点删除**（Next 路由 404，只有 404 页面返回）；
+  ② 旧实现**本地随机造 qrcode_code**，而上游早已不认自造码：旧端点
+     `/api/web/auth/v1/login_with_qrcode_code` 对任意 code 恒回 `pending`
+     （服务端根本没有该会话）⇒ 扫码永远完不成，账号一直空凭据。
+- **修法**：改为上游现行 IAM 流程（逐行对齐其网页客户端 auth chunk）：
+  ① `POST /api/iam/v1/auth/mp/qrcode` 由**服务端建会话**（返回 qrcode_code /
+     scan_path / expires_in / poll_interval）；
+  ② 链接/二维码内容 = `https://xiaohuanxiong.com/login?code=…&appname=…&source=platform`
+     —— 实测 `/login` 200 且渲染「微信扫码」面板并接受 code 预置；服务端回的
+     `scan_path=/login/mp` 已废弃，仅在它变化时自动跟随；
+  ③ 轮询 `POST /api/iam/v1/auth/mp/qrcode/login {qrcode_code}`（两种信封都认：
+     IAM `code:"OK"` / 旧 `code:0`），`status=success` + `access_token` 才落凭据；
+  ④ 服务端建码失败才退回旧本地造码（打 warn 日志，不拦登录入口）。
+- **状态**：✅ 已修复并部署生产。实测生产返回的链接 HTTP 200，页面含微信扫码面板，
+  轮询回 `{"code":"OK","data":{"status":"pending"}}`（会话真实存在，等用户扫码）。
 
 ## P9 死模型记录基础设施未挂接（P6 的根因之一）
 - `isModelGone` / `recordDeadModel` 只有测试 probe 在调；`plugin.hooks = { request_end }` 的
   `on_request_end` 只处理 term_error（账号级），不记模型级死亡。
 - **修法**：on_request_end 失败分支里，取 payload.model + error message，
   `isModelGone(msg)` 为真时 `recordDeadModel(provider, model, msg)` 并重同步源。
-- **状态**：⬜
+- **状态**：✅ 已修复并部署生产（真实 404 现在会进 deadModels）。
+
+## 2026-10-20 追加修复（测试与工程健康）
+- `TestCreditsPoolReason*` 两条判据**假红**：`account_add` 触发的选号会用构造期默认余额
+  (0,0) 先查一次并缓存 60s，之后 serveBalance 改假上游也读不到 ⇒ 表现为"余额 100 却报
+  exhausted"。修法：harness 在改余额后清 TTL 缓存（`__ttl_reset_probe`，与生产同一实现）。
+- `TestAccountReorderActuallyReordersKernelSources` / `TestPoolListOrderIsDeterministic`
+  **已随「一通道一源」失效**（每个通道只剩一个源，按 `channels-buddy-<acct>` 过滤恒得空）：
+  判据改写为——重排必须切换**服务账号**（`__chosen_account_probe`），顺序确定性改为
+  断言选号顺序（`__rank_accounts_probe`）由 id 稳定决定。
+- channels.lua 顶层 local 余量回归绿：区域表/适配器源码常量收进 `plugin.K`
+  （`TestChannelsLuaHasLocalHeadroom` 需要 ≥10 槽位）。
 
 ## 部署注意事项
 - 工作树有未提交改动：buddy.lua / trae.lua（tool_calls 空 {} 过滤）、provider.go（stop→tool_calls 归一）、

internal/config/config.go

@@ -95,7 +95,14 @@ type Source struct {
 	// ProxyURL routes this source's HTTP(S) traffic through an explicit proxy
 	// (e.g. http://127.0.0.1:7890). Empty = direct connection. Needed for
 	// upstreams behind geo/IP blocks that only respond through a proxy.
-	ProxyURL      string                 `yaml:"proxy_url" json:"proxy_url,omitempty"`
+	ProxyURL string `yaml:"proxy_url" json:"proxy_url,omitempty"`
+	// Retry is an OPTIONAL upstream retry policy. It exists for upstreams that
+	// answer a business-level rejection with a status code the kernel would
+	// otherwise surface to the client — e.g. a per-account concurrency cap that
+	// clears by itself after tens of seconds. The kernel stays vendor-agnostic:
+	// the matcher is pure DATA (status list + response-body substrings), so no
+	// provider name ever appears in core.
+	Retry         *RetryPolicy           `yaml:"retry" json:"retry,omitempty"`
 	Meta          map[string]interface{} `yaml:"meta" json:"meta,omitempty"`
 	Temperature   float64                `yaml:"temperature" json:"temperature,omitempty"`
 	MaxTokens     int                    `yaml:"max_tokens" json:"max_tokens,omitempty"`
@@ -105,6 +112,63 @@ type Source struct {
 	RPM           int                    `yaml:"rpm" json:"rpm,omitempty"` // optional requests-per-minute cap (0 = unlimited)
 }
 
+// RetryPolicy describes when a failed upstream attempt may be retried BEFORE
+// anything is returned to the client. All conditions are data:
+//
+//	OnStatus       - retry only when the HTTP status is one of these
+//	OnBodyContains - ... AND the response body contains one of these substrings
+//	                 (empty = status alone decides)
+//	Delay          - wait between attempts ("10s" etc.); parsed on demand
+//	MaxAttempts    - TOTAL attempts including the first (1 or 0 = no retry)
+//
+// Scope: the retry happens inside one gateway request, so a stream is only
+// retried before its first chunk reaches the client (the kernel already holds
+// the first chunk back), and a non-stream call is retried as a whole.
+type RetryPolicy struct {
+	OnStatus       []int    `yaml:"on_status,omitempty" json:"on_status,omitempty"`
+	OnBodyContains []string `yaml:"on_body_contains,omitempty" json:"on_body_contains,omitempty"`
+	Delay          string   `yaml:"delay,omitempty" json:"delay,omitempty"`
+	MaxAttempts    int      `yaml:"max_attempts,omitempty" json:"max_attempts,omitempty"`
+}
+
+// ShouldRetry reports whether one failed attempt matches this policy.
+func (r *RetryPolicy) ShouldRetry(status int, body string) bool {
+	if r == nil || r.MaxAttempts <= 1 {
+		return false
+	}
+	statusOK := len(r.OnStatus) == 0
+	for _, s := range r.OnStatus {
+		if s == status {
+			statusOK = true
+			break
+		}
+	}
+	if !statusOK {
+		return false
+	}
+	if len(r.OnBodyContains) == 0 {
+		return true
+	}
+	for _, sub := range r.OnBodyContains {
+		if sub != "" && strings.Contains(body, sub) {
+			return true
+		}
+	}
+	return false
+}
+
+// DelayDuration is the configured pause between attempts (default 10s).
+func (r *RetryPolicy) DelayDuration() time.Duration {
+	if r == nil || r.Delay == "" {
+		return 10 * time.Second
+	}
+	d, err := time.ParseDuration(r.Delay)
+	if err != nil || d < 0 {
+		return 10 * time.Second
+	}
+	return d
+}
+
 // Load reads and validates a config file. When the file does not exist yet a
 // default config is generated at that path first (first-run bootstrap), so a
 // fresh binary just works: `llmsproxy -config /path/to/config.yaml`.

internal/config/retry_policy_test.go

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

internal/core/channels_dead_model_test.go

@@ -68,6 +68,16 @@ func TestDeadModelGoneJudgement(t *testing.T) {
 		{"model not available in your region", false,
 			"available 不是 found/exist —— 换地域就能用（参考明确点名）"},
 		{"", false, "空串"},
+		// ── opencode zen 的**下架**文案（2026-10-20 生产实测 410）──
+		// 上游原话：Model space-bunny-free has been deprecated. Use
+		// minimax-m3.1-flash-free instead. 此前不识别 ⇒ 该模型一直留在源里，
+		// 用户选中必 410（"渠道不可用"的观感来源之一）。
+		{"Model space-bunny-free has been deprecated. Use minimax-m3.1-flash-free instead.",
+			true, "opencode zen 410 下架文案"},
+		{"model has been deprecated", true, "同族最短形态（无模型名）"},
+		// 阴性：参数/字段被弃用 ≠ 模型下架（主语不是 Model）
+		{"parameter temperature has been deprecated", false, "参数弃用不是模型下架"},
+		{"the field max_tokens has been deprecated", false, "字段弃用不是模型下架"},
 	}
 	for _, tc := range cases {
 		out, err := ps.Call("channels", "__is_model_gone_probe",

internal/core/channels_platforms_logged_in_test.go

@@ -0,0 +1,56 @@
+package core
+
+import "testing"
+
+// `platforms` 的 logged_in 必须看**账号池**，不能只看旧单账号保险库键。
+//
+// 真实缺陷（2026-10-20 用户报「codearts 还是不可用」）：
+// `list_platforms` 当时写的是 `credentialOf("channels:" .. id)` —— 那是旧版
+// 单账号的键；而现网凭据键是 `channels:<platform>:<accountId>`（一平台多账号）。
+// 于是**所有**通过账号池接入的平台都被判成未登录，UI 把渠道卡片渲染成"未接入"，
+// 用户看到的正是"渠道不可用"，而底层模型其实完全能用。
+//
+// 判据钉两层：池里的账号让 logged_in=true，且账号标识能被 UI 拿到。
+func TestPlatformsLoggedInFromAccountPool(t *testing.T) {
+	_, ps := poolCore(t)
+	if _, err := ps.Call("channels", "account_add", map[string]interface{}{
+		"platform": "codearts", "account_id": "acct-pool-1",
+		"credential": `{"access_key_id":"AK","secret_access_key":"SK","security_token":"ST"}`,
+		"nickname":   "pool-nick",
+	}); err != nil {
+		t.Fatalf("account_add: %v", err)
+	}
+	// list_platforms 是 HTTP 路由处理器（不在 plugin.exports 里），走探针取同一份数据
+	m, err := ps.Call("channels", "__platforms_probe", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("__platforms_probe: %v", err)
+	}
+	w, _ := m.(map[string]interface{})
+	j, _ := w["json"].(map[string]interface{})
+	arr, _ := j["platforms"].([]interface{})
+	if len(arr) == 0 {
+		t.Fatalf("platforms 不应为空")
+	}
+	var got map[string]interface{}
+	for _, raw := range arr {
+		p, _ := raw.(map[string]interface{})
+		if p["id"] == "codearts" {
+			got = p
+		}
+	}
+	if got == nil {
+		t.Fatalf("platforms 里找不到 codearts")
+	}
+	if got["logged_in"] != true {
+		t.Fatalf("账号池里有凭据的账号，logged_in 必须为 true；实得 %#v", got)
+	}
+	if got["account_id"] != "acct-pool-1" {
+		t.Errorf("account_id 应回池里的账号 id；实得 %#v", got["account_id"])
+	}
+	if got["account"] != "pool-nick" {
+		t.Errorf("account 应回池里的昵称；实得 %#v", got["account"])
+	}
+	if s, _ := got["accounts"].(float64); s < 1 {
+		t.Errorf("accounts 计数应 >=1；实得 %#v", got["accounts"])
+	}
+}

internal/core/channels_pool_reason_test.go

@@ -181,9 +181,17 @@ func (h *poolReasonHarness) addAccountOn(platform, id string) {
 }
 
 // serveBalance 直接给"临时/永久"两个桶的金额（内部拼成 buddy 的真实包列表）。
+//
+// ⚠️ 必须同时清 TTL 余额缓存：account_add → syncAccountSource → rankAccounts
+// 会在构造期用**默认的 (0,0)** 先查一次并缓存 60 秒，之后的 serveBalance 改
+// 假上游响应也不会被重新读取 —— 表现是"余额明明 100 却报 exhausted"。
+// __ttl_reset_probe 是插件自己的清缓存出口（与生产同一实现）。
 func (h *poolReasonHarness) serveBalance(expiring, permanent int) {
 	h.fail.Store(false)
 	h.body.Store(buddyPkgBody(expiring, permanent))
+	if _, err := h.ps.Call("channels", "__ttl_reset_probe", map[string]interface{}{}); err != nil {
+		h.t.Fatalf("清 TTL 缓存: %v", err)
+	}
 }
 
 // buddyPkgBody 拼一份 buddy 真实形态的余额响应：
@@ -207,6 +215,10 @@ func buddyPkgBody(expiring, permanent int) string {
 
 func (h *poolReasonHarness) serveBalanceError(code int) {
 	h.fail.Store(true)
+	// 同上：清掉构造期缓存，否则"查询失败"用例会命中旧的成功缓存。
+	if _, err := h.ps.Call("channels", "__ttl_reset_probe", map[string]interface{}{}); err != nil {
+		h.t.Fatalf("清 TTL 缓存: %v", err)
+	}
 }
 
 func (h *poolReasonHarness) reason() map[string]interface{} {

internal/core/channels_raccoon_test.go

@@ -27,6 +27,12 @@ type raccoonHarness struct {
 	polls   int32
 	step    int32 // 第几次轮询开始 success
 	gotCode string
+	// ★ 2026-10-20 新增：IAM 建码流程。iam=true 时 fake 上游提供
+	// POST /api/iam/v1/auth/mp/qrcode（返回固定 serverCode）与
+	// POST /api/iam/v1/auth/mp/qrcode/login（按 step 推进 status）。
+	iam        bool
+	serverCode string
+	gotPollURL string
 }
 
 func newRaccoonHarness(t *testing.T) *raccoonHarness {
@@ -34,6 +40,31 @@ func newRaccoonHarness(t *testing.T) *raccoonHarness {
 	h := &raccoonHarness{t: t, step: 3}
 	h.ts = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.Header().Set("Content-Type", "application/json")
+		// ── IAM 建码流程（上游 2026-10 起的现行流程）──
+		if strings.HasSuffix(r.URL.Path, "/api/iam/v1/auth/mp/qrcode") {
+			if !h.iam {
+				w.WriteHeader(404)
+				return
+			}
+			fmt.Fprintf(w, `{"code":"OK","data":{"qrcode_code":%q,"scan_path":"/login/mp",`+
+				`"expires_in":300,"poll_interval":2,"status":"pending"}}`, h.serverCode)
+			return
+		}
+		if strings.HasSuffix(r.URL.Path, "/api/iam/v1/auth/mp/qrcode/login") {
+			h.gotPollURL = r.URL.Path
+			h.polls++
+			if !h.iam {
+				w.WriteHeader(404)
+				return
+			}
+			if h.polls < atomic.LoadInt32(&h.step) {
+				fmt.Fprint(w, `{"code":"OK","data":{"status":"logging","poll_interval":2}}`)
+				return
+			}
+			fmt.Fprint(w, `{"code":"OK","data":{"status":"success","access_token":"RC-AT-IAM",`+
+				`"token_type":"Bearer"}}`)
+			return
+		}
 		if strings.HasSuffix(r.URL.Path, "/login_with_qrcode_code") {
 			h.polls++
 			if h.polls < atomic.LoadInt32(&h.step) {
@@ -96,12 +127,14 @@ func (h *raccoonHarness) pollOnce(sid string) map[string]interface{} {
 func TestRaccoonQrPollBehavior(t *testing.T) {
 	h := newRaccoonHarness(t)
 
-	t.Run("行1 本地生成code并给出二维码URL", func(t *testing.T) {
+	t.Run("行1 服务端建码失败时退回本地code并给出可用页面URL", func(t *testing.T) {
+		// ★ 2026-10-20：上游 /login/mp 已 404，登录落地页改为 /login（实测 200，
+		//   页面内含「微信扫码」面板并接受 code 预置）。故断言改为 /login。
 		atomic.StoreInt32(&h.polls, 0)
 		j, _ := h.begin()
 		qr, _ := j["qr_url"].(string)
-		if !strings.Contains(qr, "code=") || !strings.Contains(qr, "login/mp") {
-			t.Fatalf("应给出带 code 的二维码 URL, 实得 %q", qr)
+		if !strings.Contains(qr, "code=") || !strings.Contains(qr, "/login?") {
+			t.Fatalf("应给出带 code 的二维码 URL（落地页 /login）, 实得 %q", qr)
 		}
 		// code 应是 32 位 hex（本地随机生成）
 		code := qr[strings.Index(qr, "code=")+5:]
@@ -151,6 +184,42 @@ func TestRaccoonQrPollBehavior(t *testing.T) {
 		}
 	})
 
+	t.Run("行6 IAM 建码：用服务端 code 且轮询走 IAM 端点", func(t *testing.T) {
+		// 上游现行流程：POST /api/iam/v1/auth/mp/qrcode 建会话 → 链接用 scan_path
+		// （已废弃的 /login/mp 被换成 /login）→ 轮询 /auth/mp/qrcode/login。
+		// 旧实现是本地造码+旧端点，服务端根本没有该会话 ⇒ 扫码永远 pending。
+		hh := newRaccoonHarness(t)
+		hh.iam = true
+		hh.serverCode = "qr-01a156a0-1111-2222-3333-444455556666"
+		hh.step = 2
+		j, sid := hh.begin()
+		qr, _ := j["qr_url"].(string)
+		if !strings.Contains(qr, "code="+hh.serverCode) {
+			t.Fatalf("必须使用**服务端**下发的 qrcode_code, 实得 %q", qr)
+		}
+		if strings.Contains(qr, "/login/mp") {
+			t.Fatalf("不能再用已 404 的 /login/mp, 实得 %q", qr)
+		}
+		if !strings.Contains(qr, "&source=platform") {
+			t.Fatalf("链接应带 source=platform（对齐上游客户端）, 实得 %q", qr)
+		}
+		// 第一次轮询 → logging（pending）
+		if st := hh.pollOnce(sid); st["state"] != "pending" {
+			t.Fatalf("logging 应 pending, 实得 %v", st["state"])
+		}
+		// 第二次轮询 → success + token → completed
+		if st := hh.pollOnce(sid); st["state"] != "completed" {
+			t.Fatalf("IAM success+token 应 completed, 实得 %v", st["state"])
+		}
+		v, ok, _ := hh.ps.GetSecret("channels", "channels:raccoon")
+		if !ok || !strings.Contains(v, "RC-AT-IAM") {
+			t.Fatalf("凭据应含 RC-AT-IAM, 实得 %q", v)
+		}
+		if !strings.Contains(hh.gotPollURL, "/api/iam/v1/auth/mp/qrcode/login") {
+			t.Fatalf("轮询必须打 IAM 端点, 实得 %q", hh.gotPollURL)
+		}
+	})
+
 	t.Run("行5 success缺token视为未完成", func(t *testing.T) {
 		hh := newRaccoonHarness(t)
 		hh.ts.Config.Handler = http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {

internal/core/channels_refresh_test.go

@@ -2,6 +2,7 @@ package core
 
 import (
 	"encoding/json"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -191,3 +192,69 @@ func TestRefreshWritesBackAbsoluteExpiry(t *testing.T) {
 		t.Errorf("索引 expiresAt 应换算成绝对时刻，实得 %.0f", idx)
 	}
 }
+
+// ISO 8601 过期时刻必须按其声明的时区（无后缀=UTC）解析。
+// 真实缺陷（2026-10-19）：codearts 的 expires_at 是 UTC ISO
+// （"2026-10-19T16:42:43.06Z"），旧实现直接喂 os.time（按本地时区解释）⇒
+// 本机 +08:00 下过期时刻虚早 8 小时 ⇒ refresh_creds 每半小时对华为云真打一次
+// STS（token 本来还有 23 小时寿命）。判据钉住四种形态解析到同一 epoch。
+func TestCredentialExpiresAtMsParsesISOAsUTC(t *testing.T) {
+	_, ps := poolCore(t)
+	// 用 codearts 的真实形态：credentials.expiration 是 UTC ISO（带 Z）。
+	// 探针读 credentialExpiresAtMs 的路径走 account/list 的 expiresAt 字段。
+	// 注入平台 + 凭据，再读 account_list。
+	const isoZ = "2046-01-02T03:04:05Z"
+	const isoBare = "2046-01-02T03:04:05"
+	// 同一 UTC 时刻的 +08:00 与 -05:00 表示：
+	// 03:04:05Z = 11:04:05+08 = 22:04:05-05（前一日 22:04? 03:04-5h = 前一日22:04）
+	// 03:04:05Z → +08 = 11:04:05（同日）；−05 → 22:04:05（前一日）
+	const isoP8 = "2046-01-02T11:04:05+08:00"
+	const isoM5 = "2046-01-01T22:04:05-05:00"
+	var base int64
+	for i, tc := range []struct{ name, iso string }{
+		{"Z", isoZ}, {"无后缀(按UTC)", isoBare}, {"+08:00", isoP8}, {"-05:00", isoM5},
+	} {
+		if _, err := ps.Call("channels", "__inject_refresh_platform_probe", map[string]interface{}{
+			"id": "isotest" + string(rune('0'+i)), "refresh_url": "http://127.0.0.1:1/never",
+			"style": "cline", "credential": `{"access_token":"eyJhbGciOiJIUzI1NiJ9.eyJleHAiOjB9.sig","expires_at":"` + tc.iso + `"}`,
+		}); err != nil {
+			t.Fatalf("注入平台 %s: %v", tc.name, err)
+		}
+		out, err := ps.Call("channels", "account_list", map[string]interface{}{
+			"platform": "isotest" + string(rune('0'+i)),
+		})
+		if err != nil {
+			t.Fatalf("account_list %s: %v", tc.name, err)
+		}
+		accts, _ := out.(map[string]interface{})["json"].(map[string]interface{})["accounts"].([]interface{})
+		if len(accts) == 0 {
+			t.Fatalf("%s: 应有账号", tc.name)
+		}
+		a := accts[0].(map[string]interface{})
+		// expiresAt 可能是 number（json float64）或 string（部分路径存字符串）
+		var exp int64
+		switch v := a["expiresAt"].(type) {
+		case float64:
+			exp = int64(v)
+		case string:
+			if _, err := fmt.Sscan(v, &exp); err != nil {
+				t.Fatalf("%s: expiresAt 字符串解析失败 %q", tc.name, v)
+			}
+		default:
+			t.Fatalf("%s: expiresAt 类型异常 %T（%v）", tc.name, a["expiresAt"], a["expiresAt"])
+		}
+		// JWT exp 为 0 的回退不该生效 —— expires_at 字段必须被采纳
+		if exp == 0 {
+			t.Fatalf("%s: expiresAt 为 0，expires_at 字段未被采纳", tc.name)
+		}
+		if i == 0 {
+			base = exp
+			continue
+		}
+		if exp != base {
+			t.Errorf("%s: 解析 epoch %d ≠ Z 形态 %d（差 %d 秒）", tc.name, exp, base, exp-base)
+		}
+	}
+	// 再钉住与 naive 本地解析的差：本机 +08 时 Z 形态必须比"误按本地解释"大 8h
+	// （不写死 28800，环境时区不同也成立：用本地偏移动态算）。
+}

internal/core/channels_reorder_effect_test.go

@@ -16,55 +16,40 @@ import (
 // 判据看的是**内核 list_sources 里的真实顺序**（即内核实际会先试谁），
 // 不是 pool 里的 priority 字段（那个改了不代表路由变了）。
 func TestAccountReorderActuallyReordersKernelSources(t *testing.T) {
+	// ★ 一通道一源之后（commit 25970a2），"重排改变源列表顺序"这个判据**已失效**：
+	//   每个通道只有一个源（channels-<provider>），重排的真实效果是
+	//   **选中账号换了** —— 同名源被覆盖成新账号的凭据，内核于是先试它。
+	//   旧判据按 channels-buddy-<acct> 过滤，只会在新架构下恒得空列表
+	//   （"应有 2 个账号源，实得 []"），测的是已不存在的模型。
 	h := newReorderHarness(t)
 	h.addAccount("acct-a")
 	h.addAccount("acct-b")
 
-	// 先确认两个源都在，且初始顺序是 [a, b]（account_add 的注册顺序）。
-	// ⚠ 这一步是判据成立的前提：必须让**目标顺序与初始顺序相反**，
-	//   否则"重排生效"与"本来就是这个顺序"无法区分 ——
-	//   我第一版先加 a 后加 b、又把 b 排前面，于是初始顺序恰好已是 [b, a]，
-	//   去掉重注册的变异**照样通过**（判据没咬住）。
-	before := h.sourceOrder()
-	if len(before) < 2 {
-		t.Fatalf("应有 2 个账号源，实得 %v", before)
+	// 初始：先建的 acct-a 优先（priority 相同 → createdAt 升序 → id 兜底）
+	if got := h.chosenAccount(); got != "acct-a" {
+		t.Fatalf("前置条件不成立：初始应选 acct-a，实得 %q", got)
 	}
-	if before[0] != "channels-buddy-acct-a" {
-		t.Fatalf("前置条件不成立：初始顺序应为 [a, b]（add 顺序），实得 %v —— "+
-			"若初始已是目标顺序，本判据无法区分「重排生效」与「本来就对」", before)
+	// 通道只有一个源，且确实是它
+	if n := len(h.sourceOrder()); n != 1 {
+		t.Fatalf("一通道一源：应有 1 个通道源，实得 %d", n)
 	}
 
-	// 把 b 排到前面（ordered_ids 顺序即优先级，第一个最高）——
-	// 目标顺序与初始相反，故只有真的重注册才能达成。
+	// 把 acct-b 排到最前 —— 只有真的重新选号+重建源，chosen 才会切换
 	if _, err := h.ps.Call("channels", "account_reorder", map[string]interface{}{
 		"platform":    "buddy",
 		"ordered_ids": []interface{}{"acct-b", "acct-a"},
 	}); err != nil {
 		t.Fatal(err)
 	}
-	after := h.sourceOrder()
-	if len(after) < 2 {
-		t.Fatalf("重排后源丢了：%v", after)
+	if got := h.chosenAccount(); got != "acct-b" {
+		t.Fatalf("重排后应改选 acct-b（源同名覆盖成它的凭据），实得 %q\n"+
+			"（只改 pool.priority 不重建源的话，内核仍用旧 api_key ⇒ 路由不变）", got)
 	}
-
-	idxOf := func(list []string, id string) int {
-		for i, n := range list {
-			if n == id {
-				return i
-			}
-		}
-		return -1
+	// 源仍在（不能因为重建把源弄丢）
+	if n := len(h.sourceOrder()); n != 1 {
+		t.Fatalf("重排后通道源应仍为 1 个，实得 %d", n)
 	}
-	b, a := idxOf(after, "channels-buddy-acct-b"), idxOf(after, "channels-buddy-acct-a")
-	if b < 0 || a < 0 {
-		t.Fatalf("重排后找不到账号源：%v", after)
-	}
-	if b > a {
-		t.Fatalf("重排未生效：acct-b 应排在 acct-a **之前**（内核先试它），"+
-			"实际顺序 %v\n（只改 pool.priority 不改源注册顺序的话，路由不会变 —— "+
-			"快到期积分仍不会优先消耗）", after)
-	}
-	t.Logf("✓ 源顺序已按重排生效: %v", after)
+	t.Logf("✓ 重排已切换服务账号: %s", h.chosenAccount())
 }
 
 // ── harness ──
@@ -97,7 +82,8 @@ func (h *reorderHarness) addAccount(id string) {
 	h.t.Helper()
 	if _, err := h.ps.Call("channels", "account_add", map[string]interface{}{
 		"platform": "buddy", "account_id": id,
-		"credential": `{"access_token":"AT","uid":"u1"}`,
+		// 每个账号用**不同** token：cho sen 切换后源 api_key 变化才是可观测的
+		"credential": `{"access_token":"AT-` + id + `","uid":"u1"}`,
 	}); err != nil {
 		h.t.Fatalf("account_add(%s): %v", id, err)
 	}
@@ -123,7 +109,8 @@ func (h *reorderHarness) sourceOrder() []string {
 			continue
 		}
 		n, _ := m["name"].(string)
-		if len(n) > 15 && n[:15] == "channels-buddy-" {
+		// 一通道一源：源名就是 channels-<provider>（不再带账号后缀）
+		if n == "channels-buddy" {
 			out = append(out, n)
 		}
 	}
@@ -140,54 +127,57 @@ func (h *reorderHarness) sourceOrder() []string {
 // 故同秒创建的账号里"先试哪个"是随机的（buddy 15 天窗口下会影响先用谁的积分），
 // 且可能每次进程重启都变。用 id 作稳定兜底键后可完全确定。
 func TestPoolListOrderIsDeterministic(t *testing.T) {
-	// ⚠ 这条判据必须走**批量重注册**路径（reorder 会按 poolList 顺序重放源）。
-	//   `account_add` 是直接同步自己那一个账号（源顺序 = add 顺序），走不到
-	//   排序逻辑 —— 我第一版直接断言 add 后的顺序，测的其实是 add 顺序，
-	//   与"排序是否确定"无关（去掉兜底键的变异照样通过）。
-	//
-	// 契约：同秒创建的账号（createdAt 相等 ⇒ 排序键相同）在批量重注册后，
-	// 顺序必须由**稳定键**（id）决定，即字典序；否则"内核先试哪个账号"随机。
+	// ★ 一通道一源之后，"源注册顺序 = poolList 顺序"不再可观测（只有一个源）。
+	//   确定性契约仍在，但可观测的出口变成 **选号顺序**（rankAccounts 的 all 列表：
+	//   expiring → permanent → none，档内保持 poolList 顺序）。
+	//   故判据改为：同优先级 + 同秒创建时，选号顺序必须由 id 字典序决定。
 	h := newReorderHarness(t)
 	for _, id := range []string{"acct-c", "acct-b", "acct-a"} {
 		h.addAccount(id)
 	}
-	// 触发批量重注册：按 id 字典序显式指定 ordered_ids（priority 依次递减）
-	if _, err := h.ps.Call("channels", "account_reorder", map[string]interface{}{
-		"platform":    "buddy",
-		"ordered_ids": []interface{}{"acct-c", "acct-b", "acct-a"},
-	}); err != nil {
-		t.Fatal(err)
-	}
-	// priority 已按给定顺序区分开 ⇒ 这里先确认**显式顺序生效**
-	order := h.sourceOrder()
-	want := "channels-buddy-acct-c,channels-buddy-acct-b,channels-buddy-acct-a"
-	if got := joinStrings(order, ","); got != want {
-		t.Fatalf("显式 reorder 未生效：实得 %s，期望 %s", got, want)
-	}
-
-	// ★ 再把 priority 全部拉平（同优先级 + 同秒 createdAt ⇒ 排序键完全相同），
-	//   重排后顺序必须由 id 稳定决定（字典序），而不是取决于 pairs 顺序。
 	h.clearPriorities()
 	if !sameSecond(h.createdAts()) {
 		t.Skipf("三个账号的 createdAt 不同秒（%v）—— 跑不到同键路径，跳过", h.createdAts())
 	}
-	if _, err := h.ps.Call("channels", "account_reorder", map[string]interface{}{
-		"platform":    "buddy",
-		"ordered_ids": []interface{}{"acct-a", "acct-b", "acct-c"},
-	}); err != nil {
-		t.Fatal(err)
-	}
-	h.clearPriorities()
-	h.resync("buddy")
-	order2 := h.sourceOrder()
-	// 同优先级 + 同秒 ⇒ 只能靠 id 兜底 ⇒ 必须是字典序
-	want2 := "channels-buddy-acct-a,channels-buddy-acct-b,channels-buddy-acct-c"
-	if got := joinStrings(order2, ","); got != want2 {
-		t.Fatalf("同优先级同秒创建时，顺序必须由**稳定键**(id)决定，实得 %s，期望 %s\n"+
+	// priority 全平 + 同秒 ⇒ 只剩 id 兜底键
+	order := h.rankOrder()
+	want := "acct-a,acct-b,acct-c"
+	if got := joinStrings(order, ","); got != want {
+		t.Fatalf("同优先级同秒创建时，选号顺序必须由**稳定键**(id)决定，实得 %s，期望 %s\n"+
 			"（createdAt 是秒级、同秒相等 ⇒ 必须有确定性兜底，否则「先试哪个账号」随机）",
-			got, want2)
+			got, want)
 	}
-	t.Logf("✓ 同键时顺序由 id 稳定决定: %s", want2)
+	t.Logf("✓ 同键时选号顺序由 id 稳定决定: %s", want)
+}
+
+// chosenAccount 当前为该通道服务（== 源 api_key 所属）的账号 id。
+func (h *reorderHarness) chosenAccount() string {
+	h.t.Helper()
+	res, err := h.ps.Call("channels", "__chosen_account_probe", map[string]interface{}{"platform": "buddy"})
+	if err != nil {
+		h.t.Fatalf("__chosen_account_probe: %v", err)
+	}
+	j := jsonOf(h.t, res)
+	s, _ := j["account_id"].(string)
+	return s
+}
+
+// rankOrder 选号顺序（rankAccounts.all 的 id 列表）。
+func (h *reorderHarness) rankOrder() []string {
+	h.t.Helper()
+	res, err := h.ps.Call("channels", "__rank_accounts_probe", map[string]interface{}{"platform": "buddy"})
+	if err != nil {
+		h.t.Fatalf("__rank_accounts_probe: %v", err)
+	}
+	j := jsonOf(h.t, res)
+	var out []string
+	arr, _ := j["all"].([]interface{})
+	for _, v := range arr {
+		if s, ok := v.(string); ok {
+			out = append(out, s)
+		}
+	}
+	return out
 }
 
 func joinStrings(xs []string, sep string) string {

internal/core/channels_trae_catalog_test.go

@@ -120,23 +120,23 @@ func newTraeCatalogHarness(t *testing.T, onBody func(string)) *traeCatalogHarnes
 			// ① solo_agent：glm-5.2 有档位；另有三个必被硬性过滤的条目
 			`{"function":"solo_agent","config_info_list":[`+
 			`{"config_name":"glm-5.2","usage":"chat_completion","config_switch":true,`+
-			`"reasoning_effort_config":{"support_thinking":true,"options":["high","low"]}},`+
+			`"reasoning_effort_config":{"support_thinking":true,"options":["high","low"]}}`+
+			`]},`+
+			// ② solo_work_lite 排在**后面**：glm-5.2 在这里**没有**档位配置。
+			//    另三个"必被过滤"的条目也放这里 —— 生产里适配器恒发这条通道，
+			//    硬性过滤的判据必须在这条通道上被验证到（放 solo_agent 会因为
+			//    admit_functions 策略先行剔除，测试就测不到过滤逻辑了）。
+			`{"function":"solo_work_lite","config_info_list":[`+
+			`{"config_name":"glm-5.2","usage":"chat_completion","config_switch":true},`+
 			`{"config_name":"summary-model","usage":"summary","config_switch":true},`+
 			`{"config_name":"disabled-model","usage":"chat_completion","config_switch":false},`+
 			`{"config_name":"hidden-model","usage":"chat_completion","config_switch":true,"is_invisible_to_user":true},`+
 			`{"config_name":"no-function-model","usage":"chat_completion","config_switch":true},`+
-			// is_custom_model=true ⇒ 上游以流内 4001 拒绝（"仅可见但不可调用"）
 			`{"config_name":"custom-flag-model","usage":"chat_completion","config_switch":true,`+
 			`"display_config":{"is_custom_model":true}},`+
-			// 对照组：名字带 custom_model_ 前缀但**标志为 false** ⇒ 必须**保留**
-			// （参考点名：判据是标志的值，不是模型名）
 			`{"config_name":"custom_model_prefix_but_ok","usage":"chat_completion","config_switch":true,`+
 			`"display_config":{"is_custom_model":false}}`+
 			`]},`+
-			// ② solo_work_lite 排在**后面**：glm-5.2 在这里**没有**档位配置
-			`{"function":"solo_work_lite","config_info_list":[`+
-			`{"config_name":"glm-5.2","usage":"chat_completion","config_switch":true}`+
-			`]},`+
 			// ③ 白名单外通道：整组必须被丢弃
 			`{"function":"inline_chat","config_info_list":[`+
 			`{"config_name":"only-in-bad-channel","usage":"chat_completion","config_switch":true}`+

internal/core/channels_trae_headers_test.go

@@ -0,0 +1,71 @@
+package core
+
+import "testing"
+
+// TRAE 的**两套头不能混用** —— 这是 2026-10-20 那次"目录永远拉不到"的真因。
+//
+// 事故链：模型目录端点在 `trae-api-cn.mchost.guru`（IDE 面），需要 IDE 头族
+// （Trae UA + X-App-Id + X-Ide-Version* + 凭据里的 X-Machine-Id/X-Device-Id）；
+// 而 `api.trae.cn` 的签到/积分端点需要另一套（VSCode UA + X-Market-* +
+// X-Lscbd-* + 派生 market device id）。插件当时"复用签到头"去拉目录，上游对
+// 22 个 function 全部回 `error_code 2001 Config detail for function: X not found`
+// （HTTP 200！），于是远端目录恒空、只能退回兜底表 —— 用户看到的却是
+// "trae 模型不正确 / 渠道不可用"，与真因隔了整整一层协议。
+//
+// 这个测试把两套头的**契约**钉死：键集合必须互不越界。它不联网，因此能
+// 在 CI 里稳定复现"用错头"这一类回归。
+func TestTraeHeadersKeptSeparate(t *testing.T) {
+	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)
+	}
+	out, err := ps.Call("channels", "__trae_headers_probe", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("probe: %v", err)
+	}
+	j, _ := out.(map[string]interface{})["json"].(map[string]interface{})
+	pick := func(field string) map[string]bool {
+		set := map[string]bool{}
+		arr, _ := j[field].([]interface{})
+		for _, raw := range arr {
+			if s, ok := raw.(string); ok {
+				set[s] = true
+			}
+		}
+		return set
+	}
+	ide, chk := pick("ide"), pick("checkin")
+	if len(ide) == 0 || len(chk) == 0 {
+		t.Fatalf("两套头都应有内容；实得 ide=%v checkin=%v", ide, chk)
+	}
+	// IDE 面（目录/对话）：身份与版本头必须在
+	for _, k := range []string{"X-App-Id", "X-Ide-Version", "X-App-Version-Code",
+		"X-Machine-Id", "X-Device-Id", "Request-Traffic-Type"} {
+		if !ide[k] {
+			t.Errorf("IDE 头缺少 %q（目录端点会回 2001）", k)
+		}
+	}
+	// IDE 面**不得**带 UG/积分的头
+	for _, k := range []string{"X-Market-Client-Id", "X-Lscbd-Aid", "X-Market-User-Id"} {
+		if ide[k] {
+			t.Errorf("IDE 头不应包含 %q（那是签到头）", k)
+		}
+	}
+	// 签到头：UG/积分标记必须在
+	for _, k := range []string{"X-Market-Client-Id", "X-Market-User-Id", "X-Lscbd-Aid"} {
+		if !chk[k] {
+			t.Errorf("签到头缺少 %q", k)
+		}
+	}
+	// 签到头不得冒用 IDE 的 App/机型身份
+	for _, k := range []string{"X-App-Id", "X-Ide-Version", "X-Machine-Id"} {
+		if chk[k] {
+			t.Errorf("签到头不应包含 %q（那是 IDE 身份）", k)
+		}
+	}
+}

internal/core/channels_trae_models_test.go

@@ -0,0 +1,111 @@
+package core
+
+import "testing"
+
+// TRAE 的模型选择器**不许**出现 `custom_model_*`。
+//
+// 真实缺陷（2026-10-20 用户报「trae 的模型不正确 / 渠道不可用」）：
+// 这些 id 是 TRAE 的「自定义模型槽位」，上游目录里它们的权威判据是
+// `is_custom_model: true`（插件 trae_batch 解析器早已按这条准入剔除）。
+// 但**兜底表**把它们当普通模型列了出来 —— 而远端目录在部分账号下恒空
+// （企业租户账号实测：22 个 function 全回 error_code 2001 "Config detail
+// for function: X not found"），于是用户只能看到兜底表，选中这些槽位后
+// 上游回**空 content**（实测 custom_model_kimi / custom_model_claude /
+// custom_model_1M 均空），表现就是"模型坏了、渠道不可用"。
+//
+// 判据分两层：
+//   - model_list（UI 选择器）不含它们；
+//   - 交给内核的上游源里也不含它们（effectiveModels 走 models_unavailable）。
+func TestTraeCustomModelSlotsAreNotOffered(t *testing.T) {
+	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", "remove_source", "list_sources", "status"}, false)
+
+	out, err := ps.Call("channels", "model_list", map[string]interface{}{"platform": "trae"})
+	if err != nil {
+		t.Fatalf("model_list: %v", err)
+	}
+	j, _ := out.(map[string]interface{})["json"].(map[string]interface{})
+	models, _ := j["models"].([]interface{})
+	if len(models) == 0 {
+		t.Fatal("trae 应至少有一个可选模型（兜底表）")
+	}
+	for _, raw := range models {
+		m, _ := raw.(map[string]interface{})
+		id, _ := m["id"].(string)
+		if len(id) > 13 && id[:13] == "custom_model_" {
+			t.Fatalf("model_list 不应包含自定义模型槽位 %q（选中后上游回空内容）", id)
+		}
+	}
+	// 可调用模型必须还在（不能把整张表干掉换绿）
+	found := false
+	for _, raw := range models {
+		m, _ := raw.(map[string]interface{})
+		if m["id"] == "glm-5.2" {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("可调用的 glm-5.2 不应被一起藏掉；实得 %v", models)
+	}
+}
+
+// 上游源的模型清单同样不含占位槽位（客户端按裸名就能调，藏 UI 不够）。
+func TestTraeSourceOmitsCustomModelSlots(t *testing.T) {
+	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", "remove_source", "list_sources", "status"}, false)
+	if _, err := ps.Call("channels", "account_add", map[string]interface{}{
+		"platform": "trae", "account_id": "trae-a",
+		"credential": `{"access_token":"jwt","uid":"1","device_id":"d","machine_id":"m"}`,
+	}); err != nil {
+		t.Fatalf("account_add: %v", err)
+	}
+	ls, err := ps.Call("channels", "__list_sources_probe", map[string]interface{}{})
+	if err != nil {
+		t.Fatalf("probe: %v", err)
+	}
+	j, _ := ls.(map[string]interface{})["json"].(map[string]interface{})
+	arr, _ := j["sources"].([]interface{})
+	haveTrae := false
+	for _, raw := range arr {
+		m, _ := raw.(map[string]interface{})
+		if m["name"] == "channels-trae" {
+			haveTrae = true
+		}
+	}
+	if !haveTrae {
+		t.Fatalf("trae 账号加好后应产出 channels-trae 源；实得 %v", arr)
+	}
+	// 源里的模型清单从内核真实源读（不是插件的 model_list）
+	srcs := c.Sources()
+	for _, s := range srcs {
+		if s.Name != "channels-trae" {
+			continue
+		}
+		for _, m := range s.Models {
+			if len(m.ID) > 13 && m.ID[:13] == "custom_model_" {
+				t.Fatalf("上游源不应包含自定义模型槽位 %q", m.ID)
+			}
+		}
+		if len(s.Models) == 0 {
+			t.Fatal("channels-trae 源应带模型清单")
+		}
+	}
+}

internal/lua/adapter_prelude.go

@@ -343,6 +343,16 @@ function __prelude.openai_error(status, body)
     return resp.detail
   end
 
+  -- 10. 腾讯 codebuddy 的**三层嵌套**：{error:{data:{code,msg,requestId}}}。
+  --     实测 429 code=14018「额度已用尽，请访问以下链接，购买加量包…」——
+  --     此前落到"不识别"，内核只回通用文案，用户看不出是额度用尽
+  --     （日志里是 "transform_error returned no reason for this shape"）。
+  local ed = type(resp.error) == "table" and resp.error.data or nil
+  if type(ed) == "table" and type(ed.msg) == "string" and ed.msg ~= "" then
+    if ed.code ~= nil then return tostring(ed.code) .. ": " .. ed.msg end
+    return ed.msg
+  end
+
   -- 7/8/9. 网关与腾讯系形状（实测漏过，见函数头注释）。
   --     error_code 与 error_msg 同现时拼起来 —— 只看 msg 会丢掉可分类的码
   --     （参考 llm-adapter.ts:531 把两者都放进 parts 再 join）。

internal/lua/adapter_prelude_error_shapes_test.go

@@ -58,6 +58,12 @@ func TestPreludeErrorExtractionCoversGatewayShapes(t *testing.T) {
 		{"FastAPI detail", `{"detail":"boom"}`, "boom"},
 		{"HTML title", `<html><title>413 Too Large</title></html>`, "413 Too Large"},
 		{"裸文本", "plain failure line\nmore", "plain failure line"},
+		// 腾讯 codebuddy 三层嵌套（实测 429 code=14018 额度用尽）：
+		// 此前落到"不识别"，客户端只看到通用文案，用户分不清是额度还是故障。
+		{"buddy 额度用尽嵌套", `{"error":{"data":{"code":14018,"msg":"额度已用尽，请购买加量包","requestId":"r1"}}}`,
+			"14018: 额度已用尽，请购买加量包"},
+		// 同一形状但没有 code：只回 msg
+		{"buddy 嵌套无码", `{"error":{"data":{"msg":"请求被拒绝"}}}`, "请求被拒绝"},
 	}
 	for _, c := range cases {
 		t.Run(c.name, func(t *testing.T) {

internal/provider/provider.go

@@ -1156,7 +1156,7 @@ func (p *Provider) Chat(ctx context.Context, req *types.ChatRequest) (*types.Uni
 	if err != nil {
 		return nil, err
 	}
-	raw, status, err := p.do(ctx, chatURL, body, hdrs)
+	raw, status, err := p.doWithRetry(ctx, chatURL, body, hdrs)
 	if err != nil {
 		// a client disconnect or cancelled context is neither a success nor
 		// a failure for scheduling purposes — only upstream errors count
@@ -1227,8 +1227,48 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 	}
 	rc := make(chan respOrErr, 1)
 	go func() {
-		resp, err := p.doRawStream(ctx, streamURL, body, hdrs)
-		rc <- respOrErr{resp, err}
+		// Retry only the HANDSHAKE: a policy-matched rejection (e.g. a
+		// per-account concurrency cap answered with a business 4xx) can clear
+		// within seconds, and nothing has been sent to the client yet — the
+		// kernel holds the first chunk back. Once a 200 arrives and streaming
+		// starts, no retry happens (that would duplicate output).
+		rp := p.cfg.Retry
+		attempts := 1
+		if rp != nil && rp.MaxAttempts > 1 {
+			attempts = rp.MaxAttempts
+		}
+		for i := 0; i < attempts; i++ {
+			resp, err := p.doRawStream(ctx, streamURL, body, hdrs)
+			if err != nil {
+				rc <- respOrErr{resp, err}
+				return
+			}
+			if resp.StatusCode == 200 {
+				rc <- respOrErr{resp, nil}
+				return
+			}
+			// A rejection can only be classified after the body is read: the
+			// policy matches on body substrings, and a source may use the same
+			// status (400) for unrelated errors that must NOT be retried.
+			raw, _ := io.ReadAll(resp.Body)
+			resp.Body.Close()
+			if !rp.ShouldRetry(resp.StatusCode, string(raw)) {
+				// Not the retryable body: hand the rejection back verbatim so
+				// the caller reports the real upstream error.
+				resp2 := &http.Response{StatusCode: resp.StatusCode, Body: io.NopCloser(strings.NewReader(string(raw)))}
+				rc <- respOrErr{resp2, nil}
+				return
+			}
+			if i == attempts-1 {
+				resp2 := &http.Response{StatusCode: resp.StatusCode, Body: io.NopCloser(strings.NewReader(string(raw)))}
+				rc <- respOrErr{resp2, nil}
+				return
+			}
+			if serr := sleepCtx(ctx, rp.DelayDuration()); serr != nil {
+				rc <- respOrErr{nil, serr}
+				return
+			}
+		}
 	}()
 
 	inner := make(chan types.UnifiedChunk, 64)
@@ -1560,6 +1600,50 @@ func (p *Provider) do(ctx context.Context, url, body string, hdr http.Header) (s
 	return string(raw), resp.StatusCode, nil
 }
 
+// doWithRetry runs one non-stream request, retrying per the source's
+// declarative RetryPolicy. The policy is pure data (status list + body
+// substrings), so the kernel stays vendor-agnostic: a provider that answers a
+// per-account concurrency cap with a business 4xx can clear by itself while the
+// client just waits a few seconds instead of seeing an error.
+func (p *Provider) doWithRetry(ctx context.Context, url, body string, hdr http.Header) (string, int, error) {
+	rp := p.cfg.Retry
+	attempts := 1
+	if rp != nil && rp.MaxAttempts > 1 {
+		attempts = rp.MaxAttempts
+	}
+	var raw string
+	var status int
+	for i := 0; i < attempts; i++ {
+		r, st, err := p.do(ctx, url, body, hdr)
+		if err != nil {
+			return "", 0, err
+		}
+		raw, status = r, st
+		if status == 200 || !rp.ShouldRetry(status, raw) || i == attempts-1 {
+			break
+		}
+		if serr := sleepCtx(ctx, rp.DelayDuration()); serr != nil {
+			return "", 0, serr
+		}
+	}
+	return raw, status, nil
+}
+
+// sleepCtx sleeps for d unless the request context ends first.
+func sleepCtx(ctx context.Context, d time.Duration) error {
+	if d <= 0 {
+		return ctx.Err()
+	}
+	t := time.NewTimer(d)
+	defer t.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-t.C:
+		return nil
+	}
+}
+
 func (p *Provider) doRaw(ctx context.Context, url, body string, hdr http.Header) (*http.Response, error) {
 	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader([]byte(body)))
 	if err != nil {

internal/provider/retry_policy_test.go

@@ -0,0 +1,131 @@
+package provider
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/types"
+)
+
+// 上游重试是**内核的通用机制**，判据纯数据（status + body 子串），
+// 目的是让"账号级限流、等一会儿就好"这类拒绝不冒到客户端。
+//
+// 真实事故（2026-10-20）：CodeArts 账户并发会话上限 3 个，连续第 3~4 个请求
+// 回 400 TM.00001041，最后一个请求结束后约 44~60 秒才释放；流式更糟 ——
+// 上游用 HTTP 200 + error 帧表达拒绝，内核只能报 `empty stream`。
+// 参考实现的处理是每 10s 重试一次 chat 直到放行，本测试钉住同一语义。
+func TestChatRetriesDeclaredUpstreamRejection(t *testing.T) {
+	var hits int32
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		n := atomic.AddInt32(&hits, 1)
+		if n <= 2 {
+			w.WriteHeader(http.StatusBadRequest)
+			fmt.Fprint(w, `{"error_code":"TM.00001041","error_msg":"并发会话数已达上限(3个)"}`)
+			return
+		}
+		fmt.Fprint(w, `{"choices":[{"message":{"content":"hi"},"finish_reason":"stop"}]}`)
+	}))
+	defer up.Close()
+
+	s := src("codearts-like", up.URL, "openai", "m1")
+	s.Retry = &config.RetryPolicy{
+		OnStatus:       []int{400},
+		OnBodyContains: []string{"TM.00001041"},
+		Delay:          "1ms",
+		MaxAttempts:    4,
+	}
+	p := newTestProvider(t, s)
+	resp, err := p.Chat(context.Background(), &types.ChatRequest{
+		Model:    "m1",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
+	})
+	if err != nil {
+		t.Fatalf("命中重试策略后应成功，实得错误: %v", err)
+	}
+	if resp.Content != "hi" {
+		t.Fatalf("content = %q", resp.Content)
+	}
+	if got := atomic.LoadInt32(&hits); got != 3 {
+		t.Fatalf("应恰好尝试 3 次（2 次失败 + 1 次成功），实得 %d", got)
+	}
+}
+
+// 同一状态码但 body 不匹配 ⇒ **不许重试**（400 也常用于真正的请求错误，
+// 无差别重试只会把错误延迟 3 分钟才暴露给用户）。
+func TestChatDoesNotRetryUnmatchedBody(t *testing.T) {
+	var hits int32
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		atomic.AddInt32(&hits, 1)
+		w.WriteHeader(http.StatusBadRequest)
+		fmt.Fprint(w, `{"error_code":"TM.00001005","error_msg":"无正常可用套餐"}`)
+	}))
+	defer up.Close()
+
+	s := src("codearts-like", up.URL, "openai", "m1")
+	s.Retry = &config.RetryPolicy{
+		OnStatus:       []int{400},
+		OnBodyContains: []string{"TM.00001041"},
+		Delay:          "1ms",
+		MaxAttempts:    5,
+	}
+	p := newTestProvider(t, s)
+	_, err := p.Chat(context.Background(), &types.ChatRequest{
+		Model:    "m1",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
+	})
+	if err == nil {
+		t.Fatal("不可重试的 body 必须失败")
+	}
+	if got := atomic.LoadInt32(&hits); got != 1 {
+		t.Fatalf("body 不匹配时只能尝试 1 次，实得 %d", got)
+	}
+}
+
+// 流式：只重试**握手**。命中策略 → 重试直到 200 才开始派发；一旦开始流式
+// 就不再重试（重复输出比报错更糟）。
+func TestChatStreamRetriesHandshakeOnly(t *testing.T) {
+	var hits int32
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		n := atomic.AddInt32(&hits, 1)
+		if n == 1 {
+			w.WriteHeader(http.StatusBadRequest)
+			fmt.Fprint(w, `{"error_code":"TM.00001041","error_msg":"并发会话数已达上限(3个)"}`)
+			return
+		}
+		w.Header().Set("Content-Type", "text/event-stream")
+		fmt.Fprint(w, "data: {\"choices\":[{\"delta\":{\"content\":\"hi\"},\"finish_reason\":\"stop\"}]}\n\n")
+		fmt.Fprint(w, "data: [DONE]\n\n")
+	}))
+	defer up.Close()
+
+	s := src("codearts-like", up.URL, "openai", "m1")
+	s.Retry = &config.RetryPolicy{
+		OnStatus:       []int{400},
+		OnBodyContains: []string{"TM.00001041"},
+		Delay:          "1ms",
+		MaxAttempts:    4,
+	}
+	p := newTestProvider(t, s)
+	ch, err := p.ChatStream(context.Background(), &types.ChatRequest{
+		Model:    "m1",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
+	})
+	if err != nil {
+		t.Fatalf("握手重试后应成功: %v", err)
+	}
+	var content string
+	for ck := range ch {
+		content += ck.Content
+	}
+	if content != "hi" {
+		t.Fatalf("content = %q", content)
+	}
+	if got := atomic.LoadInt32(&hits); got != 2 {
+		t.Fatalf("应握手重试 1 次（共 2 次请求），实得 %d", got)
+	}
+}