diff --git a/apps/desktop/src/main/maker-host/__tests__/autoReviewBudget.test.ts b/apps/desktop/src/main/maker-host/__tests__/autoReviewBudget.test.ts
index ed0ba5cb030..78e866e103f 100644
--- a/apps/desktop/src/main/maker-host/__tests__/autoReviewBudget.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/autoReviewBudget.test.ts
@@ -3,9 +3,12 @@ import { describe, expect, it } from 'vitest';
 import type { CatalogModel } from '@cindy/model-providers';
 
 import {
+  AUTO_REVIEW_COMPACT_MAX_TOKENS,
+  AUTO_REVIEW_COMPACT_TIMEOUT_MS,
   findCatalogModel,
   modelCanSuppressReasoning,
   resolveAutoReviewBudget,
+  resolveAutoReviewGatewayModel,
 } from '../auto-review-budget.js';
 
 const model = (over: Partial<CatalogModel> = {}): CatalogModel => ({
@@ -44,8 +47,8 @@ describe('modelCanSuppressReasoning', () => {
 describe('resolveAutoReviewBudget', () => {
   it('keeps the compact budget for models that can turn reasoning down', () => {
     const budget = resolveAutoReviewBudget(model({ efforts: ['low', 'high'] }));
-    expect(budget.maxTokens).toBe(384);
-    expect(budget.timeoutMs).toBe(12_000);
+    expect(budget.maxTokens).toBe(AUTO_REVIEW_COMPACT_MAX_TOKENS);
+    expect(budget.timeoutMs).toBe(AUTO_REVIEW_COMPACT_TIMEOUT_MS);
     expect(budget.reasoningEffort).toBe('low');
   });
 
@@ -87,6 +90,44 @@ describe('resolveAutoReviewBudget', () => {
   });
 });
 
+describe('resolveAutoReviewGatewayModel', () => {
+  const providers = (codexModels: CatalogModel[]) => [
+    { id: 'xd', models: { codex: codexModels } },
+  ];
+
+  it('picks a suppressible gpt-class model and skips DeepSeek', () => {
+    const picked = resolveAutoReviewGatewayModel(providers([
+      model({ id: 'deepseek/deepseek-v4-pro', efforts: ['high', 'max'] }),
+      model({ id: 'moonshotai/kimi-k2.6', efforts: [] }),
+      model({ id: 'codex/gpt-5.4-mini', efforts: ['low', 'high'] }),
+    ]));
+    expect(picked?.id).toBe('codex/gpt-5.4-mini');
+  });
+
+  it('prefers budget-class mini/nano gpt over frontier gpt', () => {
+    const picked = resolveAutoReviewGatewayModel(providers([
+      model({ id: 'openai/gpt-5.6-luna', efforts: ['low', 'high'] }),
+      model({ id: 'codex/gpt-5.4-mini', efforts: ['low', 'high'] }),
+    ]));
+    expect(picked?.id).toBe('codex/gpt-5.4-mini');
+  });
+
+  it('falls back to any suppressible non-DeepSeek model when no gpt class exists', () => {
+    const picked = resolveAutoReviewGatewayModel(providers([
+      model({ id: 'deepseek/deepseek-v4-flash', efforts: ['high', 'max'] }),
+      model({ id: 'moonshotai/kimi-k2.6', efforts: [] }),
+    ]));
+    expect(picked?.id).toBe('moonshotai/kimi-k2.6');
+  });
+
+  it('returns undefined when xd only has forced-reasoning models', () => {
+    expect(resolveAutoReviewGatewayModel(providers([
+      model({ id: 'deepseek/deepseek-v4-pro', efforts: ['high', 'max'] }),
+    ]))).toBeUndefined();
+    expect(resolveAutoReviewGatewayModel([])).toBeUndefined();
+  });
+});
+
 describe('findCatalogModel', () => {
   const providers = [
     { id: 'xd', models: { 'claude-code': [model({ id: 'shared' })], pi: [model({ id: 'shared' })] } },
@@ -110,7 +151,7 @@ describe('findCatalogModel', () => {
 
   it('never borrows another provider capability when a provider is named', () => {
     // 回归 PR #2474 review:同一个模型 id 在两家目录下能力不同时,跨家借用会把
-    // 强制思考的路由误判成"能关思考",于是又拿回��凑额度 —— 正是本 PR 要修的故障。
+    // 强制思考的路由误判成"能关思考",于是又拿回紧凑额度 —— 正是本 PR 要修的故障。
     const crossProvider = [
       { id: 'xd', models: { 'claude-code': [model({ id: 'dual', efforts: ['high'] })] } },
       { id: 'other', models: { 'claude-code': [model({ id: 'dual', efforts: ['low', 'high'] })] } },
diff --git a/apps/desktop/src/main/maker-host/auto-review-budget.ts b/apps/desktop/src/main/maker-host/auto-review-budget.ts
index a81b3e3124c..fb195fcf110 100644
--- a/apps/desktop/src/main/maker-host/auto-review-budget.ts
+++ b/apps/desktop/src/main/maker-host/auto-review-budget.ts
@@ -16,8 +16,10 @@
 import type { AgentKind, CatalogModel } from '@cindy/model-providers';
 
 /** 能关思考的模型:够写一个 JSON 裁决即可。 */
-const COMPACT_MAX_TOKENS = 384;
-const COMPACT_TIMEOUT_MS = 12_000;
+export const AUTO_REVIEW_COMPACT_MAX_TOKENS = 384;
+export const AUTO_REVIEW_COMPACT_TIMEOUT_MS = 12_000;
+const COMPACT_MAX_TOKENS = AUTO_REVIEW_COMPACT_MAX_TOKENS;
+const COMPACT_TIMEOUT_MS = AUTO_REVIEW_COMPACT_TIMEOUT_MS;
 
 /**
  * 强制思考的模型:思考段 + 结论段都要装得下。
@@ -104,6 +106,31 @@ export function resolveAutoReviewBudget(model: CatalogModel | undefined): AutoRe
   };
 }
 
+/**
+ * Auto-review 灰区分类器的网关模型选型。
+ *
+ * 硬编码模型 id 不可行:xd 目录是 model-access server 按账号下发的(运行时才知道
+ * 有哪些模型,#2174)。选型规则:xd 的 codex 模型里挑「能关思考且非 DeepSeek」的,
+ * 优先 gpt 系 —— 分类器要短 JSON,强制思考模型会 malformed。
+ */
+export function resolveAutoReviewGatewayModel(
+  providers: ReadonlyArray<{
+    id: string;
+    models: Partial<Record<AgentKind, CatalogModel[]>>;
+  }>,
+): CatalogModel | undefined {
+  const xd = providers.find((item) => item.id === 'xd');
+  const models = xd?.models.codex ?? [];
+  const candidates = models.filter((m) => {
+    if (m.id.toLowerCase().includes('deepseek')) return false;
+    return modelCanSuppressReasoning(m);
+  });
+  const gptModels = candidates.filter((m) => m.id.toLowerCase().includes('gpt'));
+  // 优先 budget 档(mini/nano):分类器是高热路径,每个灰区动作都付一次模型钱。
+  const budgetClass = gptModels.find((m) => /mini|nano/.test(m.id.toLowerCase()));
+  return budgetClass ?? gptModels[0] ?? candidates[0];
+}
+
 /** 从当前目录里查一个 (供应商, agent, 模型) 的目录条目。 */
 export function findCatalogModel(
   providers: ReadonlyArray<{
diff --git a/apps/desktop/src/main/maker-host/index.ts b/apps/desktop/src/main/maker-host/index.ts
index a25c1fc4b81..086aad99775 100644
--- a/apps/desktop/src/main/maker-host/index.ts
+++ b/apps/desktop/src/main/maker-host/index.ts
@@ -16,7 +16,6 @@ import {
   ClaudeCodeAgent,
   CodexAgent,
   configureDefaultImageResizer,
-  type AutoReviewRequest,
   type McpProvider,
 } from '@cindy/maker-core';
 import {
@@ -110,7 +109,14 @@ import {
 import { resolveRemoteClaudeRoute } from './remote-claude-route.js';
 import { claudeSubagentUsageBridge } from './claude-subagent-usage-bridge.js';
 import { createAutoPermissionReviewer } from './auto-permission-reviewer.js';
-import { findCatalogModel, resolveAutoReviewBudget } from './auto-review-budget.js';
+import {
+  AUTO_REVIEW_COMPACT_MAX_TOKENS,
+  AUTO_REVIEW_COMPACT_TIMEOUT_MS,
+  findCatalogModel,
+  modelCanSuppressReasoning,
+  resolveAutoReviewBudget,
+  resolveAutoReviewGatewayModel,
+} from './auto-review-budget.js';
 import { requestUtilityText } from '../utility-model/oneShotCandidates.js';
 import { accountProviderReadinessBarrier } from './account-provider-readiness-barrier.js';
 import { hasClaudeAiOAuth } from './claude-credentials-store.js';
@@ -235,20 +241,6 @@ type RemoteCcQuery = Awaited<
 
 let _maker: Maker | null = null;
 
-/**
- * 本次审核请求的额度。按模型能力自适应:能关思考的走紧凑档,强制思考的
- * (以及目录里查不到的)给足思考+结论的空间 —— 固定 384 token 会让 DeepSeek
- * 这类模型正文恒为空。详见 auto-review-budget.ts。
- */
-function autoReviewBudgetFor(request: AutoReviewRequest) {
-  return resolveAutoReviewBudget(findCatalogModel(
-    getActiveCatalog().providers,
-    request.providerId,
-    request.agentKind,
-    request.model,
-  ));
-}
-
 let providerAccessRuntimeRefreshListener: (() => void) | null = null;
 
 /** Register the bootstrap-owned runtime reconciliation that follows provider access changes. */
@@ -258,21 +250,56 @@ export function setProviderAccessRuntimeRefreshListener(listener: (() => void) |
 
 const reviewAutoPermissionAction = createAutoPermissionReviewer({
   logger: desktopMakerLogger,
-  // 重试守卫必须按同一份额度计时,否则放宽额度的那一档会被自己的守卫切断。
-  resolveRequestTimeoutMs: (request) => autoReviewBudgetFor(request).timeoutMs,
+  // 三档候选都被约束为「能关思考」,额度恒为紧凑 12s,守卫无需再按 request 动态取。
+  resolveRequestTimeoutMs: () => AUTO_REVIEW_COMPACT_TIMEOUT_MS,
   requestText: async (request, prompt) => {
     const maker = _maker;
     if (!maker) return null;
-    const budget = autoReviewBudgetFor(request);
-    const result = await requestUtilityText(maker, prompt, {
-      providerId: request.providerId?.trim() || undefined,
-      agentKind: request.agentKind,
-      model: request.model,
-      maxTokens: budget.maxTokens,
-      timeoutMs: budget.timeoutMs,
-      ...(budget.reasoningEffort ? { reasoningEffort: budget.reasoningEffort } : {}),
+    // 1) xd 网关:按账号下发目录动态选「能关思考、非 DeepSeek」的 gpt 系模型。
+    // 额度与 effort 随选中模型算(如 efforts:[] 的 Kimi 不传 low,避免 400)。
+    const gatewayModel = resolveAutoReviewGatewayModel(getActiveCatalog().providers);
+    if (gatewayModel) {
+      const budget = resolveAutoReviewBudget(gatewayModel);
+      const result = await requestUtilityText(maker, prompt, {
+        providerId: 'xd',
+        agentKind: 'codex',
+        model: gatewayModel.id,
+        maxTokens: budget.maxTokens,
+        timeoutMs: budget.timeoutMs,
+        ...(budget.reasoningEffort ? { reasoningEffort: budget.reasoningEffort } : {}),
+      });
+      if (result.ok) return result.text;
+    }
+    // 2) ChatGPT 订阅:没有网关模型的账号回落订阅档。
+    const subscription = await requestUtilityText(maker, prompt, {
+      pinnedProfileId: 'codex-gpt-5.4-mini',
+      maxTokens: AUTO_REVIEW_COMPACT_MAX_TOKENS,
+      timeoutMs: AUTO_REVIEW_COMPACT_TIMEOUT_MS,
+      reasoningEffort: 'low',
     });
-    return result.ok ? result.text : null;
+    if (subscription.ok) return subscription.text;
+    // 3) 会话模型兜底:保留 #1227 的「跟会话 provider/model」能力,只加一条
+    // 门槛——会话模型必须能关思考。DeepSeek 等强制思考模型不再当分类器
+    // (#2174),其余自定义 provider(如 taptap + codex/gpt-5.6-terra)照旧可用。
+    const sessionModel = findCatalogModel(
+      getActiveCatalog().providers,
+      request.providerId,
+      request.agentKind,
+      request.model,
+    );
+    if (sessionModel && modelCanSuppressReasoning(sessionModel)) {
+      const budget = resolveAutoReviewBudget(sessionModel);
+      const result = await requestUtilityText(maker, prompt, {
+        providerId: request.providerId?.trim() || undefined,
+        agentKind: request.agentKind,
+        model: request.model,
+        maxTokens: budget.maxTokens,
+        timeoutMs: budget.timeoutMs,
+        ...(budget.reasoningEffort ? { reasoningEffort: budget.reasoningEffort } : {}),
+      });
+      if (result.ok) return result.text;
+    }
+    return null;
   },
 });
 
diff --git a/apps/desktop/src/main/utility-model/__tests__/oneShotCandidates.test.ts b/apps/desktop/src/main/utility-model/__tests__/oneShotCandidates.test.ts
index 7bfac0bfedf..c5b8b451b94 100644
--- a/apps/desktop/src/main/utility-model/__tests__/oneShotCandidates.test.ts
+++ b/apps/desktop/src/main/utility-model/__tests__/oneShotCandidates.test.ts
@@ -1000,6 +1000,44 @@ describe('utility one-shot candidates', () => {
     expect(body.max_tokens).toBe(64_000);
   });
 
+  it('routes Auto-review gateway codex/ model through xd /v1/responses', async () => {
+    activeCatalog.mockReturnValue({
+      providers: [{
+        id: 'xd',
+        name: 'XD',
+        source: 'builtin',
+        agents: ['codex'],
+        auth: { method: 'api-key' },
+        routing: {
+          codex: { upstream: 'https://xd.example/v1', authStrategy: 'gateway-key' },
+        },
+        models: { codex: [{ id: 'codex/gpt-5.4-mini', name: 'Mini', contextWindow: 272_000 }] },
+      }],
+    } as never);
+    readKey.mockReturnValue('xd-key');
+    fetchMock.mockResolvedValueOnce({
+      ok: true,
+      text: async () => 'data: {"type":"response.output_text.delta","delta":"{\\"verdict\\":\\"allow\\"}"}\ndata: [DONE]\n',
+    } as never);
+
+    const result = await requestUtilityText(makerMock(false), 'classify', {
+      providerId: 'xd',
+      agentKind: 'codex',
+      model: 'codex/gpt-5.4-mini',
+      maxTokens: 384,
+      reasoningEffort: 'low',
+    });
+
+    expect(result).toMatchObject({ ok: true, providerId: 'xd', model: 'codex/gpt-5.4-mini' });
+    expect(fetchMock).toHaveBeenCalledWith('https://gateway.test.invalid/v1/responses', expect.anything());
+    const body = JSON.parse(String(fetchMock.mock.calls[0]?.[1]?.body));
+    expect(body).toMatchObject({
+      model: 'codex/gpt-5.4-mini',
+      max_output_tokens: 384,
+      reasoning: { effort: 'low' },
+    });
+  });
+
   it('缺省不传 maxTokens 时,Anthropic wire 用模型目录 maxOutput 兜底(协议必填,非宿主上限)', async () => {
     activeCatalog.mockReturnValue({
       providers: [{
diff --git a/apps/desktop/src/main/utility-model/oneShotCandidates.ts b/apps/desktop/src/main/utility-model/oneShotCandidates.ts
index 3462bcc59be..9fb04e47c98 100644
--- a/apps/desktop/src/main/utility-model/oneShotCandidates.ts
+++ b/apps/desktop/src/main/utility-model/oneShotCandidates.ts
@@ -588,20 +588,23 @@ async function requestBuiltinProviderText(
     const baseUrl = effectiveXdGatewayBaseUrl().trim();
     if (!apiKey) return { ok: false, reason: 'no_candidate', attempts: [skippedAttempt(profile, 'api_key_missing')] };
     if (!baseUrl) return { ok: false, reason: 'no_candidate', attempts: [skippedAttempt(profile, 'endpoint_missing')] };
+    // Nova / Codex 折扣档(codex/…)在网关上走 Responses,chat-completions 会 400。
+    const useResponses = input.model.startsWith('codex/');
     return executeCandidates([{
       providerId: input.provider.id,
       model: input.model,
-      transport: 'litellm-chat-completions',
-      profile,
+      transport: useResponses ? 'codex-responses' : 'litellm-chat-completions',
+      profile: useResponses ? { ...profile, transport: 'codex-responses' } : profile,
       execute: (text, requestOpts) => requestProviderHttpText({
-        wire: 'chat-completions',
-        endpoint: joinProxyPath(baseUrl, '/v1/chat/completions'),
+        wire: useResponses ? 'responses' : 'chat-completions',
+        endpoint: joinProxyPath(baseUrl, useResponses ? '/v1/responses' : '/v1/chat/completions'),
         headers: { Authorization: `Bearer ${apiKey}` },
         model: input.model,
         prompt: text,
         maxTokens: requestOpts?.maxTokens ?? input.maxTokens,
         timeoutMs: requestOpts?.timeoutMs ?? input.timeoutMs,
         reasoningEffort: requestOpts?.reasoningEffort ?? input.reasoningEffort,
+        retryWithMinimalBodyOnInvalidRequest: useResponses,
       }),
     }], prompt, [], input);
   }