diff --git a/apps/ai-game-creator-shell/src/features/project-workspace/ConversationModelSelect.tsx b/apps/ai-game-creator-shell/src/features/project-workspace/ConversationModelSelect.tsx
index 1bc4c9ea4..324f83aad 100644
--- a/apps/ai-game-creator-shell/src/features/project-workspace/ConversationModelSelect.tsx
+++ b/apps/ai-game-creator-shell/src/features/project-workspace/ConversationModelSelect.tsx
@@ -14,6 +14,8 @@ import type {
   ClientLlmModel,
   ClientLlmModelCatalog,
 } from '../../services/clientApi';
+import { ClientAuthRequestError } from '../../services/clientApi';
+import { ClientHttpTimeoutError } from '../../services/clientHttp';
 import {
   cachedLlmModelCatalog,
   refreshLlmModelCatalog,
@@ -27,6 +29,14 @@ export type ConversationModelSelectHandle = {
 /** 客户端配置读取/写回失败：与「模型目录加载失败」区分，避免误导提示。 */
 class ModelSelectionConfigError extends Error {}
 
+function modelCatalogErrorMessage(error: unknown) {
+  if (error instanceof ClientHttpTimeoutError)
+    return '模型列表请求超时，请重试';
+  if (error instanceof ClientAuthRequestError && error.status)
+    return `模型列表加载失败（HTTP ${error.status}）`;
+  return '模型列表加载失败';
+}
+
 export function ConversationModelSelect({
   className,
   disabled,
@@ -51,6 +61,7 @@ export function ConversationModelSelect({
   const [busy, setBusy] = useState(!initialCatalog);
   const [error, setError] = useState('');
   const [notice, setNotice] = useState('');
+  const [manualRefreshBusy, setManualRefreshBusy] = useState(false);
   const [open, setOpen] = useState(false);
   const containerRef = useRef<HTMLDivElement | null>(null);
   const appliedRevisionRef = useRef<number | null>(
@@ -180,7 +191,11 @@ export function ConversationModelSelect({
   );
 
   const syncCatalog = useCallback(
-    async (showBusy: boolean) => {
+    async (showBusy: boolean, manualRefresh = false) => {
+      if (manualRefresh && mountedRef.current) {
+        setManualRefreshBusy(true);
+        setNotice('正在刷新模型列表');
+      }
       const busyToken = showBusy
         ? ++busyTokenRef.current
         : busyTokenRef.current;
@@ -195,12 +210,17 @@ export function ConversationModelSelect({
       try {
         let catalog: ClientLlmModelCatalog;
         let usingCachedCatalog = false;
+        let catalogError: unknown = null;
         try {
           catalog = await refreshLlmModelCatalog();
-        } catch {
+        } catch (error) {
+          catalogError = error;
           const cached = cachedLlmModelCatalog();
           if (!cached) {
-            if (mountedRef.current) setError('模型列表加载失败');
+            if (mountedRef.current) {
+              setError(modelCatalogErrorMessage(error));
+              setNotice('');
+            }
             markReady(false);
             return false;
           }
@@ -209,10 +229,14 @@ export function ConversationModelSelect({
         }
         const ready = await applyCatalog(catalog, showBusy, epochAtRequest);
         if (usingCachedCatalog && mountedRef.current)
-          setError('模型列表加载失败');
+          setError(modelCatalogErrorMessage(catalogError));
+        if (manualRefresh && mountedRef.current) {
+          setNotice(usingCachedCatalog ? '' : '模型列表已刷新');
+        }
         return ready;
       } catch (error) {
         if (mountedRef.current) {
+          setNotice('');
           setError(
             error instanceof ModelSelectionConfigError
               ? error.message
@@ -230,6 +254,7 @@ export function ConversationModelSelect({
         ) {
           setBusy(false);
         }
+        if (manualRefresh && mountedRef.current) setManualRefreshBusy(false);
       }
     },
     [applyCatalog, markReady],
@@ -375,11 +400,12 @@ export function ConversationModelSelect({
             type="button"
             className="conversation-model-menu-refresh"
             aria-label="刷新模型列表"
-            disabled={disabled || busy}
-            onClick={() => void syncCatalog(true)}
+            disabled={disabled || busy || manualRefreshBusy}
+            aria-busy={manualRefreshBusy}
+            onClick={() => void syncCatalog(true, true)}
           >
             <RefreshCcw size={13} aria-hidden="true" />
-            <span>刷新模型列表</span>
+            <span>{manualRefreshBusy ? '刷新中…' : '刷新模型列表'}</span>
           </button>
         </div>
       ) : null}
diff --git a/apps/ai-game-creator-shell/src/services/clientApi.ts b/apps/ai-game-creator-shell/src/services/clientApi.ts
index ef7af14f7..5b48072d8 100644
--- a/apps/ai-game-creator-shell/src/services/clientApi.ts
+++ b/apps/ai-game-creator-shell/src/services/clientApi.ts
@@ -8,7 +8,7 @@ import {
   ProfileWalletLedgerResponse,
   unwrapApiResponse,
 } from '../../../../packages/shared/src';
-import { fetchClientHttp } from './clientHttp';
+import { fetchClientHttp, readClientHttpResponseText } from './clientHttp';
 import { captureClientError } from './errorReporting';
 import {
   currentPlatformSessionGeneration,
@@ -48,8 +48,12 @@ export function clearStoredAuthAccessToken() {
   window.localStorage.removeItem(ACCESS_TOKEN_STORAGE_KEY);
 }
 
-async function readApiErrorMessage(response: Response, fallback: string) {
-  const text = await response.text();
+async function readApiErrorMessage(
+  response: Response,
+  fallback: string,
+  url: string,
+) {
+  const text = await readClientHttpResponseText(response, { url });
   if (!text.trim()) {
     return fallback;
   }
@@ -126,11 +130,11 @@ export async function requestClientApi<T>(
   if (!response.ok) {
     captureApiErrorStatus(url, response);
     throw new ClientAuthRequestError(
-      await readApiErrorMessage(response, fallbackMessage),
+      await readApiErrorMessage(response, fallbackMessage, url),
       { status: response.status },
     );
   }
-  const text = await response.text();
+  const text = await readClientHttpResponseText(response, { url });
   return text ? unwrapApiResponse<T>(JSON.parse(text) as T) : (null as T);
 }
 
@@ -163,7 +167,7 @@ export async function requestClientApiBytes(
   if (!response.ok) {
     captureApiErrorStatus(url, response);
     throw new ClientAuthRequestError(
-      await readApiErrorMessage(response, fallbackMessage),
+      await readApiErrorMessage(response, fallbackMessage, url),
       { status: response.status },
     );
   }
diff --git a/apps/ai-game-creator-shell/tests/clientApi.test.ts b/apps/ai-game-creator-shell/tests/clientApi.test.ts
index 7b31137a0..91d2a90c6 100644
--- a/apps/ai-game-creator-shell/tests/clientApi.test.ts
+++ b/apps/ai-game-creator-shell/tests/clientApi.test.ts
@@ -10,6 +10,12 @@ import {
   getClientAuthRefreshOperation,
   refreshClientAuthAccessToken,
 } from '../src/services/clientAuth';
+import { CLIENT_HTTP_DEFAULT_TIMEOUT_MS } from '../src/services/clientHttp';
+import {
+  cachedLlmModelCatalog,
+  refreshLlmModelCatalog,
+  resetLlmModelCatalogCacheForTest,
+} from '../src/services/llmModelCatalog';
 import {
   beginPlatformSessionTransition,
   commitAuthenticatedPlatformSession,
@@ -18,6 +24,10 @@ import {
 } from '../src/services/platformSession';
 
 vi.mock('@tauri-apps/plugin-http', () => ({ fetch: vi.fn() }));
+vi.mock(
+  '../../../packages/shared/src',
+  () => import('../../../packages/shared/src/http'),
+);
 vi.mock('../src/services/errorReporting', () => ({
   captureClientError: vi.fn(),
 }));
@@ -29,6 +39,7 @@ const json = (value: unknown, status = 200) =>
   new Response(JSON.stringify(value), { status });
 
 beforeEach(async () => {
+  resetLlmModelCatalogCacheForTest();
   resetPlatformSessionStateForTests();
   window.localStorage.clear();
   nativeInvoke.mockClear();
@@ -42,12 +53,60 @@ beforeEach(async () => {
 });
 
 afterEach(() => {
+  vi.useRealTimers();
+  resetLlmModelCatalogCacheForTest();
   resetPlatformSessionStateForTests();
   window.localStorage.clear();
   delete window.__TAURI__;
   vi.restoreAllMocks();
 });
 
+it.each([200, 503])(
+  '模型目录 HTTP %s 响应体卡住后超时，保留缓存且能再次刷新',
+  async (status) => {
+    vi.useFakeTimers();
+    const previous = { ...catalog, defaultModelId: 'quality', revision: 1 };
+    const updated = {
+      defaultModelId: 'fast',
+      models: [{ id: 'fast', displayName: '快速' }],
+      revision: 2,
+    };
+    let body!: ReadableStreamDefaultController<Uint8Array>;
+    const stalledResponse = new Response(
+      new ReadableStream<Uint8Array>({
+        start(controller) {
+          body = controller;
+        },
+      }),
+      { status },
+    );
+    const fetch = vi
+      .spyOn(globalThis, 'fetch')
+      .mockResolvedValueOnce(json(previous))
+      .mockResolvedValueOnce(stalledResponse)
+      .mockResolvedValueOnce(json(updated));
+    await expect(refreshLlmModelCatalog()).resolves.toEqual(previous);
+    let failure: unknown;
+    const pending = refreshLlmModelCatalog().catch((error: unknown) => {
+      failure = error;
+    });
+    try {
+      await vi.advanceTimersByTimeAsync(CLIENT_HTTP_DEFAULT_TIMEOUT_MS);
+      expect(failure).toMatchObject({ code: 'CLIENT_HTTP_TIMEOUT' });
+      await pending;
+      expect(cachedLlmModelCatalog()).toEqual(previous);
+      await expect(refreshLlmModelCatalog()).resolves.toEqual(updated);
+      expect(fetch).toHaveBeenCalledTimes(3);
+    } finally {
+      // 迟到的响应不能在新刷新完成后覆盖缓存，同时释放测试流。
+      body.enqueue(new TextEncoder().encode(JSON.stringify(previous)));
+      body.close();
+      await pending;
+    }
+    expect(cachedLlmModelCatalog()).toEqual(updated);
+  },
+);
+
 it('并发模型请求共享续期，并在安装 Rust 会话后使用新 token 重试', async () => {
   let refreshCalls = 0;
   let modelCalls = 0;
diff --git a/apps/ai-game-creator-shell/tests/conversationModelSelect.test.tsx b/apps/ai-game-creator-shell/tests/conversationModelSelect.test.tsx
index 4a9c7731f..9ee9e06d0 100644
--- a/apps/ai-game-creator-shell/tests/conversationModelSelect.test.tsx
+++ b/apps/ai-game-creator-shell/tests/conversationModelSelect.test.tsx
@@ -16,11 +16,36 @@ import {
   ConversationModelSelect,
   type ConversationModelSelectHandle,
 } from '../src/features/project-workspace/ConversationModelSelect';
-import { loadClientLlmModels } from '../src/services/clientApi';
+import {
+  ClientAuthRequestError,
+  type ClientLlmModelCatalog,
+  loadClientLlmModels,
+} from '../src/services/clientApi';
+import { ClientHttpTimeoutError } from '../src/services/clientHttp';
 import { resetLlmModelCatalogCacheForTest } from '../src/services/llmModelCatalog';
 
 vi.mock('../src/app/tauri', () => ({ resolveTauriInvoke: vi.fn() }));
-vi.mock('../src/services/clientApi', () => ({ loadClientLlmModels: vi.fn() }));
+const MockClientAuthRequestError = vi.hoisted(
+  () =>
+    class MockClientAuthRequestError extends Error {
+      readonly status: number | null;
+      readonly networkError: boolean;
+
+      constructor(
+        message: string,
+        options: { status?: number | null; networkError?: boolean } = {},
+      ) {
+        super(message);
+        this.status = options.status ?? null;
+        this.networkError = options.networkError ?? false;
+      }
+    },
+);
+vi.mock('../src/services/clientApi', () => ({
+  ClientAuthRequestError: MockClientAuthRequestError,
+  loadClientLlmModels: vi.fn(),
+}));
+vi.mock('@tauri-apps/plugin-http', () => ({ fetch: vi.fn() }));
 const invoke = vi.fn();
 let savedModelId = 'quality';
 let savedModelIsDefault = true;
@@ -54,6 +79,122 @@ beforeEach(() => {
 });
 afterEach(cleanup);
 
+async function renderReadyModelMenu() {
+  const onReady = vi.fn();
+  render(<ConversationModelSelect disabled={false} onReady={onReady} />);
+  await waitFor(() => expect(onReady).toHaveBeenLastCalledWith(true));
+  await act(async () => {
+    fireEvent.click(screen.getByRole('button', { name: '对话模型' }));
+  });
+  return onReady;
+}
+
+test('shows manual refresh progress immediately without clearing the selected model', async () => {
+  const onReady = await renderReadyModelMenu();
+  let resolveRefresh!: (catalog: ClientLlmModelCatalog) => void;
+  vi.mocked(loadClientLlmModels).mockImplementationOnce(
+    () =>
+      new Promise((resolve) => {
+        resolveRefresh = resolve;
+      }),
+  );
+
+  fireEvent.click(screen.getByRole('button', { name: '刷新模型列表' }));
+  expect(screen.getByRole('status').textContent).toBe('正在刷新模型列表');
+  const refreshButton = screen.getByRole('button', { name: '刷新模型列表' });
+  expect(refreshButton.textContent).toBe('刷新中…');
+  expect(refreshButton).toHaveProperty('disabled', true);
+  expect(
+    screen.getByRole('button', { name: '对话模型' }).textContent,
+  ).toContain('高质量');
+  expect(onReady).toHaveBeenLastCalledWith(false);
+
+  fireEvent.keyDown(document, { key: 'Escape' });
+  expect(screen.getByRole('status').textContent).toBe('正在刷新模型列表');
+  await act(async () => {
+    resolveRefresh({
+      defaultModelId: 'quality',
+      models: [{ id: 'quality', displayName: '高质量' }],
+      revision: 1,
+    });
+  });
+  expect(screen.getByRole('status').textContent).toBe('模型列表已刷新');
+  expect(savedModelId).toBe('quality');
+  expect(onReady).toHaveBeenLastCalledWith(true);
+});
+
+test('confirms a manual refresh even when the catalog revision is unchanged', async () => {
+  await renderReadyModelMenu();
+  expect(screen.queryByRole('status')).toBeNull();
+
+  fireEvent.click(screen.getByRole('button', { name: '刷新模型列表' }));
+  await screen.findByText('模型列表已刷新');
+  expect(loadClientLlmModels).toHaveBeenCalledTimes(3);
+  expect(
+    screen
+      .getByRole('option', { name: /高质量/ })
+      .getAttribute('aria-selected'),
+  ).toBe('true');
+  expect(screen.getByRole('option', { name: '快速' })).not.toBeNull();
+  expect(screen.getByRole('button', { name: '刷新模型列表' })).toHaveProperty(
+    'disabled',
+    false,
+  );
+});
+
+test.each([
+  [
+    'HTTP 404',
+    new ClientAuthRequestError('private server detail', { status: 404 }),
+    '模型列表加载失败（HTTP 404）',
+  ],
+  [
+    'HTTP 401',
+    new ClientAuthRequestError('private server detail', { status: 401 }),
+    '模型列表加载失败（HTTP 401）',
+  ],
+  [
+    'timeout',
+    new ClientHttpTimeoutError('https://private.example/models', 15000),
+    '模型列表请求超时，请重试',
+  ],
+  ['unknown', new Error('private server detail'), '模型列表加载失败'],
+])(
+  'reports a safe %s failure with cached models and permits retry without claiming success',
+  async (_label, failure, message) => {
+    const onReady = await renderReadyModelMenu();
+    fireEvent.click(screen.getByRole('button', { name: '刷新模型列表' }));
+    await screen.findByText('模型列表已刷新');
+
+    vi.mocked(loadClientLlmModels).mockRejectedValueOnce(failure);
+    fireEvent.click(screen.getByRole('button', { name: '刷新模型列表' }));
+    await waitFor(() =>
+      expect(screen.getByRole('alert').textContent).toBe(message),
+    );
+    expect(screen.queryByText('模型列表已刷新')).toBeNull();
+    expect(screen.queryByText('正在刷新模型列表')).toBeNull();
+    expect(document.body.textContent).not.toContain('private');
+    expect(
+      screen.getByRole('button', { name: '对话模型' }).textContent,
+    ).toContain('高质量');
+    expect(
+      screen
+        .getByRole('option', { name: /高质量/ })
+        .getAttribute('aria-selected'),
+    ).toBe('true');
+    expect(savedModelId).toBe('quality');
+    expect(onReady).toHaveBeenLastCalledWith(true);
+    expect(screen.getByRole('button', { name: '刷新模型列表' })).toHaveProperty(
+      'disabled',
+      false,
+    );
+
+    fireEvent.click(screen.getByRole('button', { name: '刷新模型列表' }));
+    await screen.findByText('模型列表已刷新');
+    expect(screen.queryByRole('alert')).toBeNull();
+  },
+);
+
 test('only displays aliases and persists selection through the native command', async () => {
   const onReady = vi.fn();
   render(<ConversationModelSelect disabled={false} onReady={onReady} />);
diff --git a/docs/project-memory/shared-memory/pitfalls.md b/docs/project-memory/shared-memory/pitfalls.md
index ebe4f1d83..00c1fa366 100644
--- a/docs/project-memory/shared-memory/pitfalls.md
+++ b/docs/project-memory/shared-memory/pitfalls.md
@@ -26,6 +26,11 @@ Direct 工具桥会 canonicalize 项目根，事件中的路径可能带 `\\?\`
 - **验证**：`apps/ai-game-creator-shell/tests/start-dev-stack.test.ts` 新增两条——「探测脚本使用 netstat 且不再出现 Get-NetTCPConnection」「命令行按 PID 缓存后随请求下发、TTL 过期即失效」；定向 vitest 55 passed。本机实测：不含 SpacetimeDB 端口的探测 368 ms（原约 22 秒）、含 SpacetimeDB 端口 3.8 秒、命中缓存 368 ms；`npm run agc:serve` 的 `starting backend stack` → `backend ready` 由约 80 秒降到 16.7 秒（其中归属校验只占 4.4 秒，其余是 SpacetimeDB + api-server 的真实启动时间）。
 - **残留**：这台机器上首次 WMI 调用本身仍是秒级（曾见 18 秒），所以「新 SpacetimeDB PID 的第一次探测」仍可能多花几秒；命令行在进程存活期内不变，TTL 只用来限制 PID 复用造成的误判窗口。
 - **关联**：`apps/ai-game-creator-shell/scripts/start-dev-stack.mjs`（`readWindowsPortOwnerIdentities`）、`apps/ai-game-creator-shell/tests/start-dev-stack.test.ts`、`apps/ai-game-creator-shell/scripts/dev-windows-process.mjs`（退出清理仍走整份 `Win32_Process` 快照，自带 1 秒缓存，不在本次范围）。
+## 2026-09-15 AGC JSON API 的响应体也必须有等待上限
+
+- `fetchClientHttp` 的超时只覆盖请求到响应头返回；随后直接等待 `response.text()` 仍可能无限挂起。模型目录共用一个在途 Promise，响应体卡住会使后续刷新复用同一挂起请求、选择器持续忙碌。
+- 成功 JSON 与错误响应体均复用 `readClientHttpResponseText` 的 15 秒上限；超时后保留最后一次有效目录并释放在途请求，手动重试重新发起请求。迟到的响应不得覆盖重试获得的新目录。
+- 排查时区分接口未挂载（404）、未授权（401）、网络或响应体超时以及刷新无变化但缺少反馈；不能仅凭客户端启动 IPC 回退警告判断刷新失败原因。
 
 ## 2026-09-14 AGC 壳 Rust 套件按「一片一 job」拆分，且分片必须自校验覆盖
 
diff --git a/docs/technical/【技术方案】AGC后台模型别名与对话选择-2026-09-05.md b/docs/technical/【技术方案】AGC后台模型别名与对话选择-2026-09-05.md
index b8e1a5335..48f119e49 100644
--- a/docs/technical/【技术方案】AGC后台模型别名与对话选择-2026-09-05.md
+++ b/docs/technical/【技术方案】AGC后台模型别名与对话选择-2026-09-05.md
@@ -7,6 +7,8 @@
 - `GET/PUT /admin/api/agc-models` 仅 owner 可用，返回完整配置；PUT 携带上次读取的 revision，冲突拒绝覆盖。
 - `GET /api/llm/models` 返回启用项的 `id/displayName`、`defaultModelId` 和目录 `revision`，不返回实际模型名、Router 目录、凭据或能力原始数据。
 - 客户端缓存最近 `revision`，在项目切换 / 对话表面挂载 / 下拉展开 / 窗口聚焦时条件刷新：`revision` 未变化不更新界面，同一时刻只保留一个在途请求，刷新失败保留上一次有效目录与本地选择。发起对话前用同一份快照校验所选模型仍启用，已停用或删除则回退默认模型并提示。
+- 手动刷新立即显示进行中状态；真实刷新成功后显示完成反馈，即使 `revision` 未变化也有反馈。失败沿用有效缓存时仍显示失败，不能报告刷新成功；HTTP 状态和超时使用可辨认的提示。
+- 模型目录与其它客户端 JSON API 的成功、失败响应体读取均复用 `readClientHttpResponseText` 的 15 秒上限；响应头已返回但响应体卡住时必须结束本次等待、释放目录在途请求并允许重试，迟到的响应不得覆盖新目录。
 - AGC Responses 请求的 `model` 是稳定目录标识。服务端按当前目录映射实际模型名；未知、停用项拒绝，不回退其它模型。旧客户端无 AGC 标记时使用后台默认项。
 - 输入框右下角选择模型，只显示别名；选择保存到客户端配置 `selectedModelId` 与 `selectedModelIsDefault`（当前选择是否来自平台默认项），从下一次请求生效。加载失败或选项停用时禁用提交并允许刷新，不显示实际 ID 作为兜底文案。
 - `selectedModelIsDefault` 为真表示选择由平台默认项驱动（首次进入、默认项变化、所选模型失效回退），后台默认项变化时客户端跟随切换并提示；用户手动选择后置为假，不再被默认项变化覆盖。
@@ -19,4 +21,5 @@
 - 目录领域校验、未知/停用模型拒绝、客户端响应不包含实际模型名。
 - 后台鉴权、持久化 revision 冲突处理；客户端选择保存后重新读取，设置保存不覆盖选择。
 - 目录 `revision` 条件刷新与并发触发去重、发送前回退默认模型、刷新失败可恢复。
+- 响应体超时保留有效缓存、再次刷新重新请求、迟到响应不覆盖新目录；手动刷新进行中、同版本成功与缓存兜底失败反馈。
 - AGC/admin-web 类型检查与定向测试、编码检查、Rust 定向检查、schema 一致性与 diff 检查。