Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(models): honor subscription context defaults for GPT-6.1 Sol (#6558)

Chris Zhang6efabf5397b1
diff --git a/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts b/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts
index 33e0627498..e1484e4371 100644
--- a/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts
@@ -821,7 +821,7 @@ describe('XD 网关权威模型清单重建', () => {
});
it.each(['google/gemini-4-argon', 'gemini-4-argon'])(
- 'Gemini 4 Argon 网关条目 %s 按公共型号补齐三引擎资料,只在原生 Pi 默认显示',
+ 'Gemini 4 Argon 网关条目 %s 采用 Server Responses 兼容配置,在 Codex / Pi 默认显示',
(id) => {
setActiveCatalog(BUNDLED_CATALOG);
setXdGatewayModels([{ id, agents: ['claude-code', 'codex', 'pi'] }]);
@@ -830,17 +830,17 @@ describe('XD 网关权威模型清单重建', () => {
expect(model, agent).toMatchObject({
name: 'Gemini 4 Argon',
group: 'google',
- nativeApi: 'google-generative-ai',
+ nativeApi: 'openai-responses',
contextWindow: 1_048_576,
supportsImageInput: true,
efforts: ['low', 'medium', 'high'],
defaultEffort: 'medium',
});
- // Claude Code / Codex 走跨 Harness bridge,默认关闭;Pi 原生支持 Google 协议。
- if (agent === 'pi') expect(model?.defaultEnabled).not.toBe(false);
- else expect(model?.defaultEnabled).toBe(false);
+ // 现有 Server 配置避开 Gateway 的 Google 原生端点限制;Claude Code bridge 仍默认关闭。
+ if (agent === 'claude-code') expect(model?.defaultEnabled).toBe(false);
+ else expect(model?.defaultEnabled).not.toBe(false);
}
- expect(resolveXdPiGatewayApi(id)).toBe('google-generative-ai');
+ expect(resolveXdPiGatewayApi(id)).toBe('openai-responses');
},
);
diff --git a/apps/desktop/src/main/maker-host/__tests__/openAiContextDefaults.test.ts b/apps/desktop/src/main/maker-host/__tests__/openAiContextDefaults.test.ts
index 22605d007d..e6a12e3217 100644
--- a/apps/desktop/src/main/maker-host/__tests__/openAiContextDefaults.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/openAiContextDefaults.test.ts
@@ -31,6 +31,49 @@ afterEach(() => {
});
describe('all OpenAI routes share the daily context default', () => {
+ // Regression: the final 272K projection discarded subscription defaults from
+ // the remote Registry. Exercise the active catalog and runtime budget lookup,
+ // including refresh and reset; existing cases only covered the 272K fallback.
+ // Unit tier: one discovered member and an in-memory catalog, no credentials.
+ it('follows subscription route defaults across harnesses and refreshes without overriding user budgets', () => {
+ const catalog = structuredClone(BUNDLED_CATALOG) as Catalog;
+ const entry = catalog.modelRegistry!.models.find(m => m.id === 'openai/gpt-6.1-sol')!;
+ const setWindow = (window: number) => {
+ entry.routes[0]!.defaults = { ...entry.routes[0]!.defaults, contextWindow: window };
+ for (const agent of ['codex', 'claude-code'] as const) {
+ entry.perAgent![agent] = { ...entry.perAgent![agent], contextWindow: window };
+ }
+ setActiveCatalog(structuredClone(catalog));
+ };
+ setDiscoveredCodexModels([{
+ id: 'gpt-6.1-sol', name: 'GPT-6.1 Sol', contextWindow: 500_000,
+ contextWindowMax: 872_000, efforts: ['medium'], defaultEffort: 'medium',
+ discoveredMetadata: { contextWindow: 500_000 },
+ }]);
+ const check = (window: number) => {
+ for (const agent of agents) {
+ const id = agent === 'codex' ? 'gpt-6.1-sol' : 'chatgpt/gpt-6.1-sol';
+ expect(row(agent, id, 'openai').contextWindow).toBe(window);
+ expect(resolveModelDefaultContextWindow(getActiveCatalog(), agent, 'openai', id)).toBe(window);
+ }
+ expect(row('codex', 'gpt-6.1-sol', 'openai').contextWindowMax).toBe(872_000);
+ };
+ setWindow(500_000);
+ check(500_000);
+ setCustomProviderConfigs([config(discovery({ slug: 'gpt-6.1-sol', context_window: 500_000 }))]);
+ for (const agent of agents) {
+ expect(row(agent, 'gpt-6.1-sol').contextWindow).toBe(272_000);
+ }
+ setWindow(400_000);
+ check(400_000);
+ setLocalCatalogOverrides({ ...EMPTY_MODEL_CATALOG_OVERRIDES, baseModels: {
+ 'openai/gpt-6.1-sol': { contextWindow: 800_000 },
+ } });
+ check(800_000);
+ setLocalCatalogOverrides(EMPTY_MODEL_CATALOG_OVERRIDES);
+ check(400_000);
+ });
+
it('imports and refreshes Sub2API without turning capacity or the default into a user edit', () => {
setActiveCatalog(BUNDLED_CATALOG);
const original = config();
diff --git a/apps/desktop/src/main/maker-host/active-catalog.ts b/apps/desktop/src/main/maker-host/active-catalog.ts
index 674cc692fc..916199ac92 100644
--- a/apps/desktop/src/main/maker-host/active-catalog.ts
+++ b/apps/desktop/src/main/maker-host/active-catalog.ts
@@ -1910,10 +1910,11 @@ function computeMerged(): Catalog {
const metadataProviderId = providerCatalogId(provider);
const rootId = metadataProviderId === 'openai'
? model.id.replace(/^chatgpt\//, '') : model.id;
- const identity = findModelRegistryRoute(
+ const matched = findModelRegistryRoute(
b.modelRegistry, metadataProviderId, rootId,
agent === 'pi' ? undefined : agent as RootAgentKind,
- )?.entry.modelRef ?? findBaseModel(b.modelRegistry, rootId)?.id;
+ );
+ const identity = matched?.entry.modelRef ?? findBaseModel(b.modelRegistry, rootId)?.id;
// Known public identities take precedence, including other vendors using
// GPT-like aliases. Only unresolved GPT/Codex/o-series IDs fall back to
// family names before the next catalog release; protocol alone never
@@ -1927,7 +1928,16 @@ function computeMerged(): Catalog {
agent as AgentKind, metadataProviderId) ||
(rootId !== model.id && hasLocalContextWindowOverride(localOverrides,
provider.id, model.id, agent as AgentKind, metadataProviderId))) return model;
- return { ...model, contextWindow: 272_000,
+ // Subscription working budgets are server-maintained per-agent defaults,
+ // separate from public model capacity. Other OpenAI connections retain
+ // the daily fallback. Pi follows the subscription Codex budget; route
+ // defaults and baseModels may contain capacity rather than a working budget.
+ const declaredWindow = isOpenAiSubscriptionProvider(provider)
+ ? matched?.entry.perAgent?.[agent === 'pi' ? 'codex' : agent as RootAgentKind]?.contextWindow
+ : undefined;
+ const contextWindow = Math.min(model.contextWindow, declaredWindow ?? 272_000);
+ if (contextWindow === model.contextWindow) return model;
+ return { ...model, contextWindow,
contextWindowMax: model.contextWindowMax ?? model.contextWindow };
}),
])),
diff --git a/apps/desktop/src/renderer/__tests__/modelDefinitionsDefaults.test.ts b/apps/desktop/src/renderer/__tests__/modelDefinitionsDefaults.test.ts
index d8c628de83..7deaecb2ff 100644
--- a/apps/desktop/src/renderer/__tests__/modelDefinitionsDefaults.test.ts
+++ b/apps/desktop/src/renderer/__tests__/modelDefinitionsDefaults.test.ts
@@ -137,12 +137,13 @@ describe('getDefaultModelForVendor', () => {
expect(md.getDefaultModelForVendor('codex').id).toBe(md.coldStartModelIdForVendor('codex'));
});
- it('cc 冷启动占位的 id 与展示名必须是 bundled 目录里排序第一的可见模型', async () => {
+ it('cc 冷启动占位的 id 与展示名必须是 bundled 目录里排序第一的可见 Claude 型号', async () => {
// 占位不是「另一份产品默认」,只是目录还没到位时的同值影子。两者漂移就会出现首帧一个
// 模型、清单到位后换成另一个的跳变。
//
- // 只锁 cc:anthropic 的动态发现清单会被 registry 用目录 sortOrder 覆盖,
- // 所以 registry 就是 cc tab 排序的权威。codex 侧不能这么锁,理由见下一条。
+ // 占位沿用 Claude 型号口径;Gateway Auto 是路由别名,不能成为订阅占位。
+ // 按公共型号身份限定参照,不因网关别名排到最前就改变冷启动默认。
+ // codex 侧不能锁相同排序,理由见下一条。
//
// 展示名一起锁住:renderer 刻意不 import BUNDLED_CATALOG(会把整份目录 JSON 打进
// bundle,为一个瞬态 label 不值得),所以 label 只能写在代码里 —— 那就由这条断言保证
@@ -151,7 +152,8 @@ describe('getDefaultModelForVendor', () => {
const { BUNDLED_CATALOG } = await import('@cindy/model-providers');
const md = await import('@/lib/modelDefinitions');
const first = Object.entries(bundledMeta(BUNDLED_CATALOG))
- .filter(([, v]) => v.agents?.includes('claude-code') && v.defaultEnabled !== false)
+ .filter(([, v]) => v.modelRef?.startsWith('anthropic/') &&
+ v.agents?.includes('claude-code') && v.defaultEnabled !== false)
.sort(
([, a], [, b]) =>
(a.sortOrder ?? Number.MAX_SAFE_INTEGER) - (b.sortOrder ?? Number.MAX_SAFE_INTEGER),
@@ -320,6 +322,7 @@ describe('newSessionDefault(专用默认种子标记)', () => {
});
interface BundledMetaEntry {
+ modelRef?: string;
agents?: readonly string[];
group?: string;
name?: string;
@@ -333,6 +336,7 @@ function bundledMeta(catalog: { modelRegistry?: ModelRegistry }): Record<string,
for (const route of entry.routes) {
if (route.providerId !== 'xd') continue;
models[route.modelId] = {
+ modelRef: entry.modelRef,
agents: route.agents,
group: entry.group,
name: entry.name,
diff --git a/apps/desktop/src/renderer/i18n/locales/en/common.json b/apps/desktop/src/renderer/i18n/locales/en/common.json
index e970c44999..5ba293fef4 100644
--- a/apps/desktop/src/renderer/i18n/locales/en/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/en/common.json
@@ -15690,6 +15690,7 @@
}
},
"modelDescriptions": {
+ "automaticRouting": "Model selection is handled by the provider.",
"longContext": "Optional long-context version; longer inputs may cost more.",
"security": "For security analysis and code review.",
"research": "For research, complex questions, coding and document creation.",
diff --git a/apps/desktop/src/renderer/i18n/locales/ja/common.json b/apps/desktop/src/renderer/i18n/locales/ja/common.json
index 6eabd2bf0c..40d9c4372d 100644
--- a/apps/desktop/src/renderer/i18n/locales/ja/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ja/common.json
@@ -15635,6 +15635,7 @@
}
},
"modelDescriptions": {
+ "automaticRouting": "モデルはプロバイダーが選択します。",
"longContext": "長い文脈を扱う追加の選択肢。入力が長いと料金が上がる場合があります。",
"security": "セキュリティ分析やコードレビューに。",
"research": "調査、複雑な問題、プログラミング、文書作成に。",
diff --git a/apps/desktop/src/renderer/i18n/locales/ko/common.json b/apps/desktop/src/renderer/i18n/locales/ko/common.json
index b1f49a3c89..36a3b46af1 100644
--- a/apps/desktop/src/renderer/i18n/locales/ko/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ko/common.json
@@ -15635,6 +15635,7 @@
}
},
"modelDescriptions": {
+ "automaticRouting": "제공자가 모델을 선택합니다.",
"longContext": "긴 문맥을 위한 선택 옵션입니다. 입력이 길면 비용이 늘어날 수 있습니다.",
"security": "보안 분석과 코드 검토에 적합합니다.",
"research": "자료 조사, 복잡한 문제, 코딩과 문서 작성에 적합합니다.",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
index c3ff85d548..2bf4dba4ff 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
@@ -15635,6 +15635,7 @@
}
},
"modelDescriptions": {
+ "automaticRouting": "由供应商选择模型。",
"longContext": "可选的长上下文版本,长输入可能产生更高费用。",
"security": "用于安全分析与代码审查。",
"research": "用于资料研究、复杂问题、编程与文档创作。",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
index 1548a1dc84..f62c5d6f19 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
@@ -15635,6 +15635,7 @@
}
},
"modelDescriptions": {
+ "automaticRouting": "由供應商選擇模型。",
"longContext": "可選的長上下文版本,長輸入可能產生更高費用。",
"security": "用於安全分析與程式碼審查。",
"research": "用於資料研究、複雜問題、程式開發與文件創作。",
diff --git a/apps/desktop/src/renderer/lib/modelDescriptions.ts b/apps/desktop/src/renderer/lib/modelDescriptions.ts
index 6097e2dc15..e88a414fc6 100644
--- a/apps/desktop/src/renderer/lib/modelDescriptions.ts
+++ b/apps/desktop/src/renderer/lib/modelDescriptions.ts
@@ -44,6 +44,10 @@ const categoryKeys: Readonly<Record<string, string>> = {
};
export function modelDescriptionKey(model: { id: string; group?: string; mode?: string }): string | undefined {
+ if ((model.id === 'auto' || model.id === 'xd/auto') &&
+ (model.mode === undefined || model.mode === 'chat' || model.mode === 'responses')) {
+ return 'automaticRouting';
+ }
// A vendor grouping like "gpt" must not turn GPT Image into a coding model.
const inferredCategory = classifyModel({ id: model.id, mode: model.mode });
const category = model.mode !== undefined || Object.hasOwn(categoryKeys, inferredCategory)
diff --git a/docs/dev-rules/model-catalog-runtime.md b/docs/dev-rules/model-catalog-runtime.md
index 56733b07c2..a4f6596514 100644
--- a/docs/dev-rules/model-catalog-runtime.md
+++ b/docs/dev-rules/model-catalog-runtime.md
@@ -128,15 +128,22 @@ wire 字段或修改远程目录优先级。Server 仍可按原 schema/revision
用户重命名改变匹配;使用稳定的模型 ID。添加系列时补齐五语资源,覆盖测试遍历本地
Registry 的全部模型及其 routes,防止只翻译当前默认启用的几款。
-### GPT 日常窗口与 Codex Chat Completions(2026-09-06 用户裁决)
+### GPT 日常窗口与 Codex Chat Completions(2026-10-20 更新)
-- 所有来源的 OpenAI 聊天模型采用至多 272,000 tokens 的日常默认窗口,包括内置订阅、XD、
+- OpenAI 聊天模型未声明订阅工作默认时采用至多 272,000 tokens 的日常默认窗口,包括内置订阅、XD、
自定义 API/Sub2API、第三方预设和组织连接,覆盖 Claude Code、Codex、Pi。
覆盖 GPT、Codex、o 系列的原始 ID、`codex/`、`openai/`、`chatgpt/` 别名,以及 Registry
明确关联到 OpenAI 公共型号的部署 ID。已解析的公共型号身份优先于名称;只有未解析身份时
才按上述系列与命名空间推断,任意私有命名空间不推断。较小模型不扩容;仅有兼容协议不推断厂商。
`contextWindowMax` 保留供应商容量;这是客户端工作默认策略,不修改服务端能力声明。
-- 显式上下文 override 仍优先,可设置 1M;恢复默认删除 override 后采用 272K。
+- OpenAI 订阅的日常窗口由 Server 既有 `routes[].defaults.contextWindow` 与
+ `perAgent[引擎].contextWindow` 下发;最终工作预算采用 perAgent,Pi 跟随 Codex 的订阅预算。活动目录以该工作
+ 默认限制已解析窗口,不扩容较小的实报,也不从 route 或 `baseModels` 的总容量推导工作默认。
+ GPT-6.1 Sol 的订阅默认为 500,000,Astra 仍为 272,000;自定义 API、XD 和第三方预设
+ 不继承订阅工作预算。用户覆盖仍最高。实现及刷新/恢复默认回归见
+ `active-catalog.ts` 与 `openAiContextDefaults.test.ts`。
+- 显式上下文 override 仍优先,可设置 1M;恢复默认删除 override 后重新读取对应订阅工作默认,
+ 未声明的路由使用 272K。
用户显式模型窗口、公共型号与连接/引擎补丁仍优先;非 OpenAI 模型和媒体模型不套用此策略。
导入、刷新、离线回退、组织同步与本地 addition 经同一活动目录出口投影,不把默认值写回发现快照或用户配置。
- Codex CLI 0.153.0 已移除原生 Chat Completions。Cindy 的既有转换路径仍可使用,
diff --git a/docs/product-rules/model-metadata-precedence.md b/docs/product-rules/model-metadata-precedence.md
index dec734e3fb..673b44c320 100644
--- a/docs/product-rules/model-metadata-precedence.md
+++ b/docs/product-rules/model-metadata-precedence.md
@@ -221,6 +221,12 @@ Server `catalog/providers.json` 是数据正本;客户端 `catalog/model-regis
复用当前目录接口与 Registry V4 协商,不增加请求或数据库表。旧 V1/V2/V3 客户端收到展开后的旧字段;不同资料的多条路由在兼容响应中拆成独立条目,保留上游 ID,额外条目使用派生目录 ID。旧协议不支持的公共引用、覆盖指令与图片字段被移除,空默认以缺省表达。
-供应商总容量与客户端工作预算分开:所有来源 OpenAI 聊天模型的至多 272K 工作默认继续保留(含第三方、自定义与组织连接,用户显式设置优先);这里的 `contextWindowMax` 指客户端 CatalogModel 投影,不是 Server 字段。Registry 资料中的 contextWindow 与合法 perAgent 工作默认分别维护;用户显式窗口覆盖仍优先。协议和实际执行能力门禁继续由运行时负责,元数据不能解锁未实现协议或伪造可售性。
+供应商总容量与客户端工作预算分开:OpenAI 订阅采用 Server Registry 的 `perAgent` 工作默认,
+Pi 跟随 Codex 的订阅预算;GPT-6.1 Sol 为 500K,Astra 仍为 272K。未声明订阅预算及其它来源
+(含第三方、自定义与组织连接)继续采用至多 272K 的工作默认,不扩大较小的已解析窗口。
+不能从 route 或公共型号的总容量推导工作预算。用户显式窗口覆盖仍最高,恢复默认重新读取当前
+订阅预算或 272K 回退。这里的 `contextWindowMax` 指客户端 CatalogModel 投影,不是 Server 字段,
+不随工作默认缩小;Registry 资料中的 contextWindow 与合法 perAgent 工作默认分别维护。
+协议和实际执行能力门禁继续由运行时负责,元数据不能解锁未实现协议或伪造可售性。
验收至少覆盖缺字段继承、供应商优先、force 修正、用户最高、显式 false/null/空数组、未知供应商公共识别、刷新与保存、旧版解析、离线与坏快照回退。模型推荐标准另见 [本地模型筛选](local-model-selection.md),仍只按能力、速度和实际运行内存筛选。
diff --git a/packages/model-providers/catalog/model-registry.json b/packages/model-providers/catalog/model-registry.json
index 931418d14e..b86c6187c8 100644
--- a/packages/model-providers/catalog/model-registry.json
+++ b/packages/model-providers/catalog/model-registry.json
@@ -1,6 +1,6 @@
{
"schemaVersion": 5,
- "updatedAt": "2026-10-14T00:00:00.000Z",
+ "updatedAt": "2026-10-20T09:39:02.203Z",
"models": [
{
"id": "openai/gpt-6.1-sol",
@@ -16,7 +16,8 @@
"codex"
],
"defaults": {
- "supportsFastMode": true
+ "supportsFastMode": true,
+ "contextWindow": 500000
},
"referencePriceGroup": "global"
}
@@ -25,7 +26,7 @@
"modelRef": "openai/gpt-6.1-sol",
"perAgent": {
"codex": {
- "contextWindow": 272000,
+ "contextWindow": 500000,
"efforts": [
"low",
"medium",
@@ -36,7 +37,7 @@
]
},
"claude-code": {
- "contextWindow": 272000,
+ "contextWindow": 500000,
"efforts": [
"low",
"medium",
@@ -1035,7 +1036,12 @@
"providerId": "xai",
"modelId": "xai/grok-4.7",
"agents": ["claude-code", "codex"],
- "referencePriceGroup": "global"
+ "referencePriceGroup": "global",
+ "forceOverrides": {
+ "contextWindow": 500000
+ },
+ "overrideReason": "SuperGrok subscription /v1/models reports context_window 256000 (the Grok CLI working window), but the subscription responses endpoint accepts prompts up to 500000 tokens and rejects larger ones with 'context window (N tokens > 500000 tokens)'.",
+ "overrideVerifiedAt": "2026-10-20"
}
],
"nativeApi": "openai-responses",
@@ -1056,7 +1062,12 @@
"providerId": "xai",
"modelId": "xai/grok-4.7-build-fast",
"agents": ["claude-code", "codex"],
- "referencePriceGroup": "global"
+ "referencePriceGroup": "global",
+ "forceOverrides": {
+ "contextWindow": 500000
+ },
+ "overrideReason": "SuperGrok subscription /v1/models reports context_window 256000 (the Grok CLI working window), but the subscription responses endpoint accepts prompts up to 500000 tokens and rejects larger ones with 'context window (N tokens > 500000 tokens)'.",
+ "overrideVerifiedAt": "2026-10-20"
}
],
"nativeApi": "openai-responses",
@@ -1072,7 +1083,12 @@
"providerId": "xai",
"modelId": "xai/grok-4.6",
"agents": ["claude-code", "codex"],
- "referencePriceGroup": "global"
+ "referencePriceGroup": "global",
+ "forceOverrides": {
+ "contextWindow": 500000
+ },
+ "overrideReason": "SuperGrok subscription /v1/models reports context_window 256000 (the Grok CLI working window), but the subscription responses endpoint accepts prompts up to 500000 tokens and rejects larger ones with 'context window (N tokens > 500000 tokens)'.",
+ "overrideVerifiedAt": "2026-10-20"
}
],
"nativeApi": "openai-responses",
@@ -1088,7 +1104,12 @@
"providerId": "xai",
"modelId": "xai/grok-4.5",
"agents": ["claude-code", "codex"],
- "referencePriceGroup": "global"
+ "referencePriceGroup": "global",
+ "forceOverrides": {
+ "contextWindow": 500000
+ },
+ "overrideReason": "SuperGrok subscription /v1/models reports context_window 256000 (the Grok CLI working window), but the subscription responses endpoint accepts prompts up to 500000 tokens and rejects larger ones with 'context window (N tokens > 500000 tokens)'.",
+ "overrideVerifiedAt": "2026-10-20"
}
],
"nativeApi": "openai-responses",
@@ -2490,9 +2511,34 @@
],
"nativeApi": "openai-completions",
"modelRef": "xiaomi/mimo-v2.6-pro-ultraspeed"
+ },
+ {
+ "id": "xd/auto",
+ "name": "Auto",
+ "status": "active",
+ "sortOrder": -1,
+ "routes": [
+ {
+ "providerId": "xd",
+ "modelId": "auto",
+ "agents": ["claude-code", "codex"]
+ }
+ ],
+ "nativeApi": "openai-responses",
+ "modelRef": "xd/auto"
}
],
"nativeApiRules": [
+ {
+ "providerId": "xd",
+ "modelIdPrefix": "google/gemini-4-argon",
+ "nativeApi": "openai-responses"
+ },
+ {
+ "providerId": "xd",
+ "modelIdPrefix": "gemini-4-argon",
+ "nativeApi": "openai-responses"
+ },
{
"providerId": "xd",
"modelIdPrefix": "google/gemini-",
@@ -7381,6 +7427,13 @@
},
"officialDocs": "https://www.alibabacloud.com/help/en/model-studio/tts-model"
}
+ },
+ {
+ "id": "xd/auto",
+ "aliases": [],
+ "defaults": {
+ "name": "Auto"
+ }
}
]
}
diff --git a/packages/model-providers/src/__tests__/gemini4ArgonCatalog.test.ts b/packages/model-providers/src/__tests__/gemini4ArgonCatalog.test.ts
index 295f863410..550598605b 100644
--- a/packages/model-providers/src/__tests__/gemini4ArgonCatalog.test.ts
+++ b/packages/model-providers/src/__tests__/gemini4ArgonCatalog.test.ts
@@ -29,14 +29,18 @@ describe("Gemini 4 Argon catalog entry", () => {
});
});
- it("stays a public model without a guessed gateway route", () => {
+ it("uses the published Gateway transport without inventing model availability", () => {
expect(
registry?.models.some((m) => m.modelRef === "google/gemini-4-argon"),
).toBe(false);
for (const id of ["google/gemini-4-argon", "gemini-4-argon"])
expect(resolveModelNativeApi(registry, "xd", id), id).toBe(
- "google-generative-ai",
+ "openai-responses",
);
+ // The Server's Argon compatibility rule is narrower than the Gemini family.
+ expect(resolveModelNativeApi(registry, "xd", "google/gemini-3.8-flash")).toBe(
+ "google-generative-ai",
+ );
});
it("switches from the launch price to the published 2026-12-01 price", () => {
diff --git a/packages/model-providers/src/__tests__/modelRegistryConsistency.test.ts b/packages/model-providers/src/__tests__/modelRegistryConsistency.test.ts
index 9a990bb10b..984e8cd195 100644
--- a/packages/model-providers/src/__tests__/modelRegistryConsistency.test.ts
+++ b/packages/model-providers/src/__tests__/modelRegistryConsistency.test.ts
@@ -171,7 +171,7 @@ describe("model registry data consistency", () => {
expect(
effectiveWindow(entry, agent),
`${entry.id}/${agent}`,
- ).toBe(272_000);
+ ).toBe(entry.id === "openai/gpt-6.1-sol" ? 500_000 : 272_000);
}
continue;
}