dsh-gpt-adapter 0.2.3 → 0.2.5

package/README.md

@@ -12,6 +12,7 @@ Install one plugin to reuse local Codex instructions, skills, enabled plugin res
 
 - Discover upstream models with `GET /models` and retain stale entries until explicit cleanup.
 - Provide the GPT reasoning preset `low / medium / high / xhigh / max / ultra` through the native Models page and conversation menu.
+- Reuse the native per-model context field for manual capacities and automatic resolution, with support for K/M input.
 - Offer independent Standard and Fast speed choices. Fast sends `service_tier: "priority"`; Ultrafast requires an explicit upstream declaration or override.
 - Preserve optional function arguments on GPT Responses requests by setting missing `strict` values to `false`.
 - Read Codex instruction files and refresh them at the next model step without repeating unchanged baselines.
@@ -22,14 +23,14 @@ Reasoning and speed presets are requests to the configured gateway, not guarante
 
 ## Compatibility and installation
 
-The native integration targets DSH `0.2.2-rc.1`, Cordis `4.1.0`, and the pinned adapter source hash. The development dependency uses Pi AI `1.3.0`. Installed third-party package files are never edited. The published npm package is `dsh-gpt-adapter@0.2.3`.
+The native integration targets DSH `0.2.2-rc.1`, Cordis `4.1.0`, and the pinned adapter source hash. The development dependency uses Pi AI `1.3.0`. Installed third-party package files are never edited. The package version is `dsh-gpt-adapter@0.2.5`.
 
 Requires Node.js `>=24.9.0` and DSH. Source builds also require pnpm.
 
 Install the published package in the web profile:
 
 ```sh
-dsh plugin --profile web add dsh-gpt-adapter@0.2.3
+dsh plugin --profile web add dsh-gpt-adapter@0.2.5
 ```
 
 Restart the DSH host after installing or upgrading. Configure GPT connections under **Settings → Models**; Codex environment reuse is enabled by default. The plugin extends the existing native settings and conversation controls.
@@ -42,7 +43,7 @@ cd dsh-gpt-adapter
 pnpm install --frozen-lockfile
 mkdir -p artifacts
 pnpm pack --pack-destination artifacts
-dsh plugin --profile web add ./artifacts/dsh-gpt-adapter-0.2.3.tgz
+dsh plugin --profile web add ./artifacts/dsh-gpt-adapter-0.2.5.tgz
 ```
 
 `pnpm pack` runs the production build.
@@ -125,6 +126,24 @@ An empty `mcpAllowlist` imports all enabled declared servers. A nonempty list fi
 
 Capability precedence is user override, explicit upstream metadata, external catalog, then preset. Catalog requests never receive the provider API key. Model-list responses may omit capability metadata. Reasoning is sent as `reasoning.effort` for Responses and `reasoning_effort` for Chat Completions; the selected protocol is preserved.
 
+### Context windows
+
+Under **Settings → Models → Custom settings → Model catalog**, edit the native **Context window** field. The plugin reuses this field and adds no separate context editor. Enter a count, `516K` (`516000` tokens), or `1M` (`1000000` tokens); clear the field and save to restore automatic resolution.
+
+Inherited model rows and provider defaults participate in edit detection and final budget validation. A failed legacy cleanup restores the previous native capacity while preserving unrelated edits.
+
+An empty native field means automatic resolution. Automatic capacities stay in the effective model information; the editable draft contains manual values only, including existing legacy overrides. An explicit value equal to the current automatic capacity is still saved as a manual value.
+
+Capacity suffixes use decimal units. Automatic resolution uses a distinguishable explicit provider default, upstream model metadata, the configured capability catalog, the shipped exact-model preset, then the native catalog/DSH fallback. Shipped presets are dated data, not a live fetch from OpenAI: `gpt-6-astra`, `gpt-6.1-sol`, `gpt-6.1-terra`, `gpt-6-luna`, and `gpt-5.4` use `1050000`; `gpt-4.1` uses `1047576`; `gpt-4o` uses `128000`. Only documented snapshot aliases share these presets. References and verification dates are in [the preset table](src/context-window.ts).
+
+Existing native per-model values, including catalog `modelOverrides`, and legacy JSON context overrides remain effective until deliberately edited. Editing the native context field clears only that model's legacy context override; clearing the field removes its manual context values while retaining provider defaults and other settings. Saving unrelated settings or discovering models does not turn automatic values into manual overrides. Identical model IDs on different providers are independent.
+
+Providers using native catalog `modelOverrides` keep their catalog configuration during synchronization. Discovery and stale cleanup update the plugin snapshot without creating an incompatible explicit `models` list or deleting native overrides.
+
+Native forms reject stale drafts when their automatic capacities change; refresh the form before retrying. Context, output-limit, and relevant provider-default edits are checked against the saved draft's output budget and native compaction headroom. An omitted output limit uses the native model catalog or provider/schema default. Invalid saves leave the previous values in place.
+
+Changes apply to the next prepared call; calls already prepared retain their captured capacity. Model information, the inference model, overflow detection, and native compaction use the same effective window. Maximum output tokens and DSH's compaction policy remain independent. A larger local window does not increase the gateway's accepted limit.
+
 Preset reasoning fields are projected into the native UI rather than persisted into native model configuration. Disabling GPT adaptation restores the native model values and schema. Speed selections are persisted per session. Synchronization failures retain the previous snapshot; absent models remain stale until explicit cleanup.
 
 GPT model snapshots and speed selections are stored under `$DSH_HOME/storages/gpt-adapter`. Codex source files and imported MCP credentials are not copied there. The DSH host owns normal session transcripts and logs.

package/README.zh-CN.md

@@ -12,6 +12,7 @@
 
 - 通过 `GET /models` 发现上游模型；失效模型保留到用户主动清理。
 - 在原生模型设置页和对话菜单中提供 `low / medium / high / xhigh / max / ultra` 六档思考预设。
+- 复用原生逐模型上下文字段，支持 K/M 输入、手动容量和自动解析。
 - 提供独立的 Standard 和 Fast 速度选择。Fast 发送 `service_tier: "priority"`；Ultrafast 由上游明确声明或用户覆盖提供。
 - 为 GPT Responses 函数工具补齐缺失的 `strict: false`，保留可选参数。
 - 读取 Codex 指令，在下一模型步骤更新；内容不变时不重复注入。
@@ -22,14 +23,14 @@
 
 ## 兼容性与安装
 
-原生集成面向 DSH `0.2.2-rc.1`、Cordis `4.1.0`，并校验适配器源码指纹。开发依赖固定 Pi AI `1.3.0`。插件不会修改已安装的第三方包文件。当前 npm 发布版本为 `dsh-gpt-adapter@0.2.3`。
+原生集成面向 DSH `0.2.2-rc.1`、Cordis `4.1.0`，并校验适配器源码指纹。开发依赖固定 Pi AI `1.3.0`。插件不会修改已安装的第三方包文件。插件版本为 `dsh-gpt-adapter@0.2.5`。
 
 需要 Node.js `>=24.9.0` 和 DSH。源码构建另外需要 pnpm。
 
 在 `web` 配置档（profile）中安装已发布的 npm 包：
 
 ```sh
-dsh plugin --profile web add dsh-gpt-adapter@0.2.3
+dsh plugin --profile web add dsh-gpt-adapter@0.2.5
 ```
 
 安装或升级后重启 DSH 宿主。在「设置 → 模型」中配置 GPT 连接；Codex 环境复用默认开启。插件扩展现有原生设置与对话控件。
@@ -42,7 +43,7 @@ cd dsh-gpt-adapter
 pnpm install --frozen-lockfile
 mkdir -p artifacts
 pnpm pack --pack-destination artifacts
-dsh plugin --profile web add ./artifacts/dsh-gpt-adapter-0.2.3.tgz
+dsh plugin --profile web add ./artifacts/dsh-gpt-adapter-0.2.5.tgz
 ```
 
 `pnpm pack` 会自动执行生产构建。
@@ -125,6 +126,24 @@ codexCompatibility:
 
 能力优先级为用户覆盖、上游明确元数据、外部能力目录、预设。能力目录请求不携带提供商 API 密钥；模型列表可以不包含能力元数据。Responses 使用 `reasoning.effort`，Chat Completions 使用 `reasoning_effort`，保留用户选定的协议。
 
+### 上下文窗口
+
+在「设置 → 模型 → 自定义设置 → 模型目录」中，直接编辑原生「上下文窗口」字段。插件复用这一字段，不再增加独立编辑入口。可输入完整词元数、`516K`（`516000`）或 `1M`（`1000000`）；清空字段并保存可恢复自动解析。
+
+编辑前后状态比较和最终预算校验会纳入继承模型及提供商默认值。旧覆盖清理失败时，恢复原生容量并保留无关修改。
+
+原生字段留空表示自动解析。自动容量保留在模型有效信息中，编辑草稿只显示手动值（包括已有旧版覆盖）；明确输入与当前自动容量相同的数值，也会保存为手动值。
+
+容量后缀使用十进制单位。自动解析依次采用可辨识的显式提供商默认值、上游模型元数据、已配置的能力目录、插件内置的精确模型预设、原生目录或 DSH 后备值。内置预设是带核验日期的数据，不会实时抓取 OpenAI 页面：`gpt-6-astra`、`gpt-6.1-sol`、`gpt-6.1-terra`、`gpt-6-luna`、`gpt-5.4` 为 `1050000`，`gpt-4.1` 为 `1047576`，`gpt-4o` 为 `128000`；仅为有文档依据的快照别名复用预设。参考页面和核验日期见[预设表](src/context-window.ts)。
+
+已有原生模型数值（包括目录模型的 `modelOverrides`）和旧版 JSON 上下文覆盖继续生效，直到用户主动编辑。编辑原生上下文字段时，仅清除该模型的旧版上下文覆盖；清空字段会删除该模型的手动上下文数值，保留提供商默认值和其他设置。保存无关设置或发现模型不会把自动值固化为手动覆盖。同名模型在不同提供商下独立设置。
+
+使用原生目录 `modelOverrides` 的提供商在同步时保留目录配置。模型发现和失效清理只更新插件快照，不会生成不兼容的显式 `models` 列表，也不会删除原生覆盖。
+
+原生表单的自动容量变化后，过期草稿会被拒绝，刷新表单再保存即可。修改上下文、最大输出或相关提供商默认值时，会校验同次保存的输出预算和原生压缩预留；未填写输出上限时采用原生模型目录或提供商、参数结构的默认值；无效数值不会覆盖已保存设置。
+
+修改对下一次准备的调用生效，已准备的调用继续使用原有容量。模型信息、推理模型、溢出判断和原生上下文压缩采用同一有效窗口；最大输出词元设置与 DSH 原生压缩策略保持独立。扩大本地窗口不会提高网关实际接受的上限。
+
 思考预设只投影到原生界面，不写入原生模型配置。关闭 GPT 自动适配后恢复原生值和参数结构。速度选择按会话持久化。同步失败保留旧快照，失效模型保留到主动清理。
 
 GPT 模型快照与速度选择保存在 `$DSH_HOME/storages/gpt-adapter`。Codex 源文件和导入的 MCP 凭证不会复制到该位置。正常会话记录和日志由 DSH 宿主管理。

package/dist/src/context-window.d.ts

@@ -0,0 +1,25 @@
+export type ContextWindowMode = 'default' | '516k' | '1m' | 'custom';
+export type ContextWindowSource = 'legacy-override' | 'native-user' | 'provider-default' | 'upstream' | 'capability-catalog' | 'official-preset' | 'native-default';
+export type ContextWindowResolution = {
+    mode: ContextWindowMode;
+    effectiveTokens: number;
+    defaultTokens: number;
+    source: ContextWindowSource;
+};
+export declare const OFFICIAL_CONTEXT_WINDOWS: Record<string, {
+    tokens: number;
+    url: string;
+    verifiedAt: string;
+}>;
+export declare function validContextWindow(value: unknown): value is number;
+export declare function contextWindowMode(tokens: number): ContextWindowMode;
+export declare function contextWindowTokens(mode: ContextWindowMode, customTokens?: number): number | undefined;
+export declare function resolveContextWindow(options: {
+    modelId: string;
+    legacy?: unknown;
+    native?: unknown;
+    providerDefault?: unknown;
+    upstream?: unknown;
+    catalog?: unknown;
+    nativeDefault?: unknown;
+}): ContextWindowResolution;

package/dist/src/host.d.ts

@@ -2,6 +2,8 @@ import { Service, type Context } from '@deepseek-ai/cordis';
 import z from '@deepseek-ai/schemastery';
 import { bindTypertRemote } from '@deepseek-ai/dsh-typert-protocol';
 import type { SessionId } from '@deepseek-ai/dsh-client-connection/client';
+import { type ContextWindowMode, type ContextWindowResolution } from './context-window.js';
+import { type ContextWindowEdit, type ContextWindowWriter } from './settings-overlay.js';
 import { type CodexCompatibilityConfig, type CodexCompatibilityRuntime } from './codex.js';
 import type { ModelCapabilities, ServiceTier } from './types.js';
 type RouteOptions = {
@@ -51,6 +53,7 @@ export declare class GptModels extends Service {
     private readonly options;
     private readonly settingsNs;
     codex?: CodexCompatibilityRuntime;
+    contextWriter: ContextWindowWriter | undefined;
     readonly typertRemote: ReturnType<typeof bindTypertRemote>;
     private readonly runners;
     private readonly snapshots;
@@ -85,12 +88,36 @@ export declare class GptModels extends Service {
         capabilities: {
             [x: string]: ModelCapabilities;
         };
+        contextWindows: {
+            [k: string]: ContextWindowResolution;
+        };
+        contextRevision: number;
         options: RouteOptions;
     };
     configure(provider: string, enabled: boolean): Promise<void>;
     updateOptions(provider: string, refreshIntervalMs: number, capabilityCatalogUrl: string, overridesJSON: string): Promise<void>;
     refresh(provider: string): Promise<void>;
     removeStale(provider: string): Promise<void>;
+    /** Internal model-list writes must use the same manual fields as native drafts. */
+    private editableModel;
+    editableContextWindow(provider: string, model: Record<string, any>): any;
+    /** Resolve raw metadata separately from merged capabilities and cached overrides. */
+    contextWindow(provider: string, modelId: string, nativeDefault?: number, profileOverride?: {
+        defaultContextWindow?: number;
+        models?: Array<{
+            id: string;
+            [key: string]: unknown;
+        }>;
+        modelOverrides?: Record<string, Record<string, unknown>>;
+    }): ContextWindowResolution;
+    /** Both native form saves and the compatibility RPC validate the same budget. */
+    validateContextWindow(provider: string, model: Record<string, any>, profile: Record<string, any>): void;
+    contextWriteRevision(): number | undefined;
+    /** Public form revisions also cover changes to projected model capacities. */
+    private modelsRevision;
+    /** Remove only legacy context fields after the matching native save commits. */
+    clearLegacyContext(edits: ContextWindowEdit[], expected?: number): Promise<void>;
+    setContextWindow(provider: string, modelId: string, mode: ContextWindowMode, customTokens?: number, expectedRevision?: number): Promise<void>;
     setSpeed(sessionId: SessionId, provider: string, modelId: string, tier: ServiceTier): Promise<void>;
     speed(sessionId: SessionId, provider: string, modelId: string): {
         selected: ServiceTier;

package/dist/src/settings-overlay.d.ts

@@ -5,9 +5,23 @@ type SettingsLike = {
     write?: (ns: string, change: (current: any, base: any, schema: any) => any, expected?: number, paths?: string[][]) => Promise<void>;
 };
 export type CapabilityResolver = (provider: string, model: JsonObject, profile?: JsonObject) => ModelCapabilities | undefined;
-export declare function projectModel(provider: string, model: JsonObject, resolve: CapabilityResolver, profile?: JsonObject): JsonObject;
+export type ContextWindowEdit = {
+    provider: string;
+    modelId: string;
+    previousTokens: number | undefined;
+};
+type ContextWrites = {
+    editableContext?(provider: string, model: JsonObject): number | undefined;
+    validate?(provider: string, model: JsonObject, profile: JsonObject): void;
+    revision(): number | undefined;
+    saved(edits: ContextWindowEdit[], revision: number | undefined): Promise<void>;
+};
+export type ContextWindowWriter = (provider: string, modelId: string, tokens: number | undefined, expected?: number) => Promise<void>;
+export declare function projectModel(provider: string, model: JsonObject, resolve: CapabilityResolver, profile?: JsonObject, includeContext?: boolean): JsonObject;
 /** A view of native configuration; the original object and its stored fields stay intact. */
 export declare function projectConfiguration(value: JsonObject | undefined, resolve: CapabilityResolver): JsonObject | undefined;
 /** Dispose returns the native settings page to its original schema and values. */
-export declare function installSettingsOverlay(settings: SettingsLike, resolve: CapabilityResolver): () => void;
+export declare function installSettingsOverlay(settings: SettingsLike, resolve: CapabilityResolver, contextWrites?: ContextWrites): (() => void) & {
+    setContextWindow: ContextWindowWriter;
+};
 export {};

package/lib/client.js

@@ -48,17 +48,21 @@ function apply(ctx) {
     const [value, setValue] = (0, import_react.useState)();
     const [error, setError] = (0, import_react.useState)("");
     const [busy, setBusy] = (0, import_react.useState)(false);
+    const requests = (0, import_react.useRef)({ sequence: 0, writing: false });
     (0, import_react.useEffect)(() => {
       let active = true;
+      requests.current.writing = false;
+      setBusy(false);
       const refresh = () => {
-        if (!provider) return;
+        if (!provider || requests.current.writing) return;
+        const sequence = ++requests.current.sequence;
         void call("status", { provider }).then((value2) => {
-          if (active) {
+          if (active && sequence === requests.current.sequence) {
             setValue(value2);
             setError("");
           }
         }).catch((error2) => {
-          if (active)
+          if (active && sequence === requests.current.sequence)
             setError(
               error2 instanceof Error ? error2.message : "\u65E0\u6CD5\u8BFB\u53D6 GPT \u8BBE\u7F6E"
             );
@@ -68,22 +72,32 @@ function apply(ctx) {
       const timer = setInterval(refresh, 3e3);
       return () => {
         active = false;
+        requests.current.sequence++;
         clearInterval(timer);
       };
     }, [provider]);
     async function act(method, args) {
-      if (busy) return;
+      if (requests.current.writing) return;
+      requests.current.writing = true;
+      const sequence = ++requests.current.sequence;
       setBusy(true);
       try {
         await call(method, args);
-        setValue(await call("status", { provider }));
+        if (sequence !== requests.current.sequence) return false;
+        const next = await call("status", { provider });
+        if (sequence !== requests.current.sequence) return false;
+        setValue(next);
         setError("");
         return true;
       } catch (error2) {
-        setError(error2 instanceof Error ? error2.message : "\u64CD\u4F5C\u5931\u8D25\uFF0C\u8BF7\u91CD\u8BD5");
+        if (sequence === requests.current.sequence)
+          setError(error2 instanceof Error ? error2.message : "\u64CD\u4F5C\u5931\u8D25\uFF0C\u8BF7\u91CD\u8BD5");
         return false;
       } finally {
-        setBusy(false);
+        if (sequence === requests.current.sequence) {
+          requests.current.writing = false;
+          setBusy(false);
+        }
       }
     }
     return { value, error, busy, act };
@@ -220,6 +234,7 @@ function apply(ctx) {
             ]
           }
         ) }),
+        /* @__PURE__ */ (0, import_jsx_runtime.jsx)("p", { className: native_settings_styles_default.advancedHint, children: "\u4E0A\u4E0B\u6587\u7A97\u53E3\u8BF7\u5728\u300C\u81EA\u5B9A\u4E49\u8BBE\u7F6E \u2192 \u6A21\u578B\u76EE\u5F55\u300D\u7684\u539F\u751F\u5B57\u6BB5\u4E2D\u4FEE\u6539\uFF0C\u652F\u6301 516K\u30011M \u6216\u81EA\u5B9A\u4E49\u6570\u503C\uFF1B\u6E05\u7A7A\u540E\u4FDD\u5B58\u53EF\u6062\u590D\u81EA\u52A8\u503C\u3002" }),
         /* @__PURE__ */ (0, import_jsx_runtime.jsx)(
           AdvancedSettings,
           {

package/lib/core.js

@@ -121,6 +121,33 @@ function capabilityForSelection(capabilities, reasoningId, serviceTier2) {
   };
 }
 
+// src/context-window.ts
+var officialModels = [
+  ["gpt-6-astra", 105e4],
+  ["gpt-6.1-sol", 105e4],
+  ["gpt-6.1-terra", 105e4],
+  ["gpt-6-luna", 105e4],
+  ["gpt-5.4", 105e4],
+  ["gpt-4.1", 1047576],
+  ["gpt-4o", 128e3],
+  ["gpt-5.4-2026-03-05", 105e4, "gpt-5.4"],
+  ["gpt-4.1-2025-04-14", 1047576, "gpt-4.1"],
+  ["gpt-4o-2024-08-06", 128e3, "gpt-4o"]
+];
+var OFFICIAL_CONTEXT_WINDOWS = Object.fromEntries(
+  officialModels.map(([id, tokens, page]) => [
+    id,
+    {
+      tokens,
+      url: `https://developers.openai.com/api/docs/models/${page ?? id}`,
+      verifiedAt: "2026-11-13"
+    }
+  ])
+);
+function validContextWindow(value) {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0;
+}
+
 // src/sync.ts
 import { createHash } from "node:crypto";
 var MAX_RESPONSE_BYTES = 4 * 1024 * 1024;
@@ -298,7 +325,7 @@ function parseCapabilityHints(row, modelId) {
     ...typeof defaultReasoningValue === "string" ? { defaultReasoningId: defaultReasoningValue } : {},
     ...hasTierSignal ? { serviceTiers: tiers } : {},
     ...defaultTier ? { defaultServiceTier: defaultTier } : {},
-    ...typeof contextWindow === "number" && contextWindow > 0 ? { contextWindow } : {},
+    ...validContextWindow(contextWindow) ? { contextWindow } : {},
     ...typeof maxOutputTokens === "number" && maxOutputTokens > 0 ? { maxOutputTokens } : {}
   };
   return Object.keys(hints).length === 0 ? void 0 : hints;

package/lib/index.js

@@ -196,6 +196,84 @@ function createRequestBridge(provider) {
   };
 }
 
+// src/context-window.ts
+var officialModels = [
+  ["gpt-6-astra", 105e4],
+  ["gpt-6.1-sol", 105e4],
+  ["gpt-6.1-terra", 105e4],
+  ["gpt-6-luna", 105e4],
+  ["gpt-5.4", 105e4],
+  ["gpt-4.1", 1047576],
+  ["gpt-4o", 128e3],
+  ["gpt-5.4-2026-03-05", 105e4, "gpt-5.4"],
+  ["gpt-4.1-2025-04-14", 1047576, "gpt-4.1"],
+  ["gpt-4o-2024-08-06", 128e3, "gpt-4o"]
+];
+var OFFICIAL_CONTEXT_WINDOWS = Object.fromEntries(
+  officialModels.map(([id, tokens, page]) => [
+    id,
+    {
+      tokens,
+      url: `https://developers.openai.com/api/docs/models/${page ?? id}`,
+      verifiedAt: "2026-11-13"
+    }
+  ])
+);
+function validContextWindow(value) {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0;
+}
+function contextWindowMode(tokens) {
+  return tokens === 516e3 ? "516k" : tokens === 1e6 ? "1m" : "custom";
+}
+function contextWindowTokens(mode, customTokens) {
+  if (!["default", "516k", "1m", "custom"].includes(mode))
+    throw new Error("\u8BF7\u9009\u62E9\u6709\u6548\u7684\u4E0A\u4E0B\u6587\u9884\u8BBE");
+  if (mode !== "custom" && customTokens !== void 0)
+    throw new Error("\u53EA\u6709\u81EA\u5B9A\u4E49\u9884\u8BBE\u63A5\u53D7\u7A97\u53E3\u6570\u503C");
+  if (mode === "custom") {
+    if (!validContextWindow(customTokens))
+      throw new Error("\u4E0A\u4E0B\u6587\u7A97\u53E3\u5FC5\u987B\u662F\u6B63\u7684\u5B89\u5168\u6574\u6570");
+    return customTokens;
+  }
+  return mode === "516k" ? 516e3 : mode === "1m" ? 1e6 : void 0;
+}
+function resolveContextWindow(options) {
+  const fallback = [
+    [options.providerDefault, "provider-default"],
+    [options.upstream, "upstream"],
+    [options.catalog, "capability-catalog"],
+    [
+      Object.hasOwn(OFFICIAL_CONTEXT_WINDOWS, options.modelId) ? OFFICIAL_CONTEXT_WINDOWS[options.modelId]?.tokens : void 0,
+      "official-preset"
+    ],
+    [options.nativeDefault, "native-default"],
+    [1e6, "native-default"]
+  ];
+  const [defaultTokens, source] = fallback.find(
+    ([tokens]) => validContextWindow(tokens)
+  );
+  for (const [tokens, manualSource] of [
+    [options.legacy, "legacy-override"],
+    [options.native, "native-user"]
+  ]) {
+    if (tokens === void 0) continue;
+    if (!validContextWindow(tokens))
+      throw new Error("\u4E0A\u4E0B\u6587\u7A97\u53E3\u5FC5\u987B\u662F\u6B63\u7684\u5B89\u5168\u6574\u6570");
+    return {
+      mode: contextWindowMode(tokens),
+      effectiveTokens: tokens,
+      defaultTokens,
+      source: manualSource
+    };
+  }
+  return {
+    mode: "default",
+    effectiveTokens: defaultTokens,
+    defaultTokens,
+    source
+  };
+}
+
 // src/sync.ts
 import { createHash } from "node:crypto";
 var MAX_RESPONSE_BYTES = 4 * 1024 * 1024;
@@ -373,7 +451,7 @@ function parseCapabilityHints(row, modelId) {
     ...typeof defaultReasoningValue === "string" ? { defaultReasoningId: defaultReasoningValue } : {},
     ...hasTierSignal ? { serviceTiers: tiers } : {},
     ...defaultTier ? { defaultServiceTier: defaultTier } : {},
-    ...typeof contextWindow === "number" && contextWindow > 0 ? { contextWindow } : {},
+    ...validContextWindow(contextWindow) ? { contextWindow } : {},
     ...typeof maxOutputTokens === "number" && maxOutputTokens > 0 ? { maxOutputTokens } : {}
   };
   return Object.keys(hints).length === 0 ? void 0 : hints;
@@ -731,10 +809,11 @@ var JsonStorage = class {
 };
 
 // src/host.ts
-import { PiAiAdapter } from "@deepseek-ai/dsh-llm-pi-ai";
+import { Config as NativeConfig, PiAiAdapter } from "@deepseek-ai/dsh-llm-pi-ai";
 import { createRequire } from "node:module";
 import { readFileSync } from "node:fs";
 import { createHash as createHash6 } from "node:crypto";
+import { getBuiltinModels } from "@earendil-works/pi-ai/providers/all";
 
 // src/native-adapter.ts
 function installNativeAdapterBridge(prototype, capture) {
@@ -759,6 +838,12 @@ function installNativeAdapterBridge(prototype, capture) {
       });
     const models = new Proxy(snapshot.models, {
       get(target, key) {
+        if (key === "getModel" && typeof target.getModel === "function")
+          return (provider, modelId) => {
+            const model = target.getModel(provider, modelId);
+            const capacity = extension.model.context;
+            return model && provider === options.provider && modelId === options.model && capacity?.contextWindow ? { ...model, contextWindow: capacity.contextWindow } : model;
+          };
         if (key === "streamSimple")
           return (model, context, settings) => {
             const call = extension.prepare(options);
@@ -853,11 +938,12 @@ function installNativeAdapterBridge(prototype, capture) {
 // src/settings-overlay.ts
 import z from "@deepseek-ai/schemastery";
 import { symbols } from "@deepseek-ai/cordis";
-function projectModel(provider, model, resolve4, profile) {
+function projectModel(provider, model, resolve4, profile, includeContext = true) {
   const capability = resolve4(provider, model, profile);
   if (!capability) return model;
   return {
     ...model,
+    ...includeContext && validContextWindow(capability.contextWindow) ? { contextWindow: capability.contextWindow } : {},
     reasoningEfforts: capability.reasoning.status === "unsupported" ? false : Object.fromEntries(
       capability.reasoning.choices.map((choice) => [
         choice.id,
@@ -903,52 +989,91 @@ function projectSchema(json, preserveNativeLevels) {
   keys.list = offered.map((level) => z.const(level));
   return schema.toJSON();
 }
-function nativeWrite(next, current, resolve4) {
+function nativeWrite(next, current, base, resolve4, contextEdits, contextWrites) {
   if (!next.providers) return next;
   return {
     ...next,
     providers: Object.fromEntries(
       Object.entries(next.providers).map(([provider, raw]) => {
         const profile = raw;
-        if (!Array.isArray(profile?.models)) return [provider, profile];
-        const previous = current.providers?.[provider]?.models ?? [];
+        const inherited = base?.providers?.[provider] ?? {};
+        const ownershipProfile = {
+          ...inherited,
+          ...current.providers?.[provider]
+        };
+        const effectiveProfile = { ...inherited, ...profile };
+        const ownsModels = Array.isArray(profile?.models);
+        const models = ownsModels ? profile.models : inherited.models?.map(
+          (model) => projectModel(provider, model, resolve4, ownershipProfile)
+        );
+        if (!Array.isArray(models)) return [provider, profile];
+        const previous = ownershipProfile.models ?? [];
+        const savedModels = models.map((model) => {
+          const original = previous.find((row) => row.id === model.id);
+          const capability = resolve4(provider, model, ownershipProfile);
+          if (!capability) return model;
+          const managed = projectModel(
+            provider,
+            model,
+            resolve4,
+            ownershipProfile
+          );
+          if (original && JSON.stringify(model.reasoningEfforts) !== JSON.stringify(managed.reasoningEfforts))
+            throw new Error(
+              "GPT \u601D\u8003\u7B49\u7EA7\u7531\u63D2\u4EF6\u9884\u8BBE\u7BA1\u7406\uFF1B\u8BF7\u5237\u65B0\u540E\u91CD\u8BD5\uFF0C\u6216\u5728\u6A21\u578B\u80FD\u529B\u8986\u76D6\u4E2D\u4FEE\u6539\uFF0C\u505C\u7528\u81EA\u52A8\u9002\u914D\u53EF\u6062\u590D\u539F\u751F\u7F16\u8F91\u3002"
+            );
+          const restored = { ...model };
+          const projectedContext = original ? projectModel(provider, original, resolve4, ownershipProfile).contextWindow : void 0;
+          const editedContext = ownsModels && (original ? model.contextWindow !== projectedContext : Object.hasOwn(model, "contextWindow"));
+          const budgetChanged = editedContext || !original || model.maxTokens !== original.maxTokens || model.maxTokens === void 0 && effectiveProfile.defaultMaxTokens !== ownershipProfile.defaultMaxTokens || model.contextWindow === void 0 && effectiveProfile.defaultContextWindow !== ownershipProfile.defaultContextWindow;
+          if (!editedContext) {
+            if (original && Object.hasOwn(original, "contextWindow"))
+              restored.contextWindow = original.contextWindow;
+            else delete restored.contextWindow;
+          } else {
+            if (model.contextWindow !== void 0 && !validContextWindow(model.contextWindow))
+              throw new Error("\u4E0A\u4E0B\u6587\u7A97\u53E3\u5FC5\u987B\u662F\u6B63\u7684\u5B89\u5168\u6574\u6570");
+            contextEdits.push({
+              provider,
+              modelId: model.id,
+              previousTokens: original?.contextWindow
+            });
+          }
+          if (budgetChanged)
+            contextWrites?.validate?.(
+              provider,
+              editedContext ? model : { ...model, contextWindow: capability.contextWindow },
+              profile
+            );
+          if (original && Object.hasOwn(original, "reasoningEfforts"))
+            restored.reasoningEfforts = structuredClone(
+              original.reasoningEfforts
+            );
+          else delete restored.reasoningEfforts;
+          return restored;
+        });
         return [
           provider,
-          {
-            ...profile,
-            models: profile.models.map((model) => {
-              const original = previous.find((row) => row.id === model.id);
-              const ownershipProfile = current.providers?.[provider] ?? profile;
-              const capability = resolve4(provider, model, ownershipProfile);
-              if (!capability) return model;
-              const managed = projectModel(
-                provider,
-                model,
-                resolve4,
-                ownershipProfile
-              );
-              if (original && JSON.stringify(model.reasoningEfforts) !== JSON.stringify(managed.reasoningEfforts))
-                throw new Error(
-                  "GPT \u601D\u8003\u7B49\u7EA7\u7531\u63D2\u4EF6\u9884\u8BBE\u7BA1\u7406\uFF1B\u8BF7\u5237\u65B0\u540E\u91CD\u8BD5\uFF0C\u6216\u5728\u6A21\u578B\u80FD\u529B\u8986\u76D6\u4E2D\u4FEE\u6539\uFF0C\u505C\u7528\u81EA\u52A8\u9002\u914D\u53EF\u6062\u590D\u539F\u751F\u7F16\u8F91\u3002"
-                );
-              const restored = { ...model };
-              if (original && Object.hasOwn(original, "reasoningEfforts"))
-                restored.reasoningEfforts = structuredClone(
-                  original.reasoningEfforts
-                );
-              else delete restored.reasoningEfforts;
-              return restored;
-            })
-          }
+          ownsModels ? { ...profile, models: savedModels } : profile
         ];
       })
     )
   };
 }
-function installSettingsOverlay(settings, resolve4) {
+function installSettingsOverlay(settings, resolve4, contextWrites) {
   settings = settings[symbols.original] ?? settings;
   const originalDescribe = settings.describe;
   const originalWrite = settings.write;
+  const editableResolve = (provider, model, profile) => {
+    const capability = resolve4(provider, model, profile);
+    if (!capability) return void 0;
+    const manual = contextWrites?.editableContext ? contextWrites.editableContext(provider, model) : model.contextWindow;
+    const projected = { ...capability };
+    if (manual === void 0) delete projected.contextWindow;
+    else projected.contextWindow = manual;
+    return projected;
+  };
+  let viewRevision;
   function describe(options) {
     return originalDescribe.call(this, options).map((row) => {
       if (row.ns !== "llm-pi-ai") return row;
@@ -960,40 +1085,140 @@ function installSettingsOverlay(settings, resolve4) {
         for (const model of profile.models ?? [])
           if (resolve4(provider, model, profile)) managed = true;
           else other = true;
-      if (!managed) return row;
+      const value = managed ? projectConfiguration(row.value, resolve4) : row.value;
+      const contexts = JSON.stringify(
+        Object.entries(value?.providers ?? {}).map(([provider, profile]) => [
+          provider,
+          profile.models?.map((model) => [
+            model.id,
+            model.contextWindow
+          ])
+        ])
+      );
+      if (typeof row.revision === "number") {
+        if (!viewRevision)
+          viewRevision = {
+            native: row.revision,
+            contexts,
+            revision: row.revision
+          };
+        else if (viewRevision.native !== row.revision || viewRevision.contexts !== contexts)
+          viewRevision = {
+            native: row.revision,
+            contexts,
+            revision: Math.max(row.revision, viewRevision.revision + 1)
+          };
+      }
       return {
         ...row,
-        schema: projectSchema(row.schema, other),
-        value: projectConfiguration(row.value, resolve4),
-        base: projectConfiguration(row.base, resolve4),
-        user: projectConfiguration(row.user, resolve4)
+        ...managed ? {
+          schema: projectSchema(row.schema, other),
+          value,
+          base: projectConfiguration(row.base, editableResolve),
+          user: projectConfiguration(row.user, editableResolve)
+        } : {},
+        ...typeof row.revision === "number" ? { revision: viewRevision.revision } : {}
       };
     });
   }
   async function write(ns, change, expected, paths) {
-    return originalWrite.call(
+    const edits = [];
+    const legacyRevision = contextWrites?.revision();
+    await originalWrite.call(
       this,
       ns,
       ns !== "llm-pi-ai" ? change : (current, base, schema) => nativeWrite(
         change(
-          projectConfiguration(current, resolve4),
-          projectConfiguration(base, resolve4),
+          projectConfiguration(structuredClone(current), editableResolve),
+          projectConfiguration(structuredClone(base), editableResolve),
           schema
         ),
         current,
-        resolve4
+        base,
+        editableResolve,
+        edits,
+        contextWrites
       ),
       expected,
       paths
     );
+    await finishContextWrite.call(this, ns, edits, legacyRevision, paths);
+  }
+  async function finishContextWrite(ns, edits, legacyRevision, paths) {
+    if (ns !== "llm-pi-ai" || !edits.length || !contextWrites) return;
+    const committedRevision = describe.call(this).find((row) => row.ns === ns)?.revision;
+    try {
+      await contextWrites.saved(edits, legacyRevision);
+    } catch (error) {
+      try {
+        await originalWrite.call(
+          this,
+          ns,
+          (current) => {
+            const restored = structuredClone(current);
+            for (const edit of edits) {
+              const row = restored.providers?.[edit.provider]?.models?.find(
+                (model) => model.id === edit.modelId
+              );
+              if (!row) continue;
+              if (edit.previousTokens === void 0) delete row.contextWindow;
+              else row.contextWindow = edit.previousTokens;
+            }
+            return restored;
+          },
+          committedRevision,
+          paths
+        );
+      } catch (rollbackError) {
+        throw new AggregateError(
+          [error, rollbackError],
+          "\u4E0A\u4E0B\u6587\u4FDD\u5B58\u672A\u5B8C\u6210\uFF0C\u6062\u590D\u5931\u8D25\uFF1B\u8BF7\u5237\u65B0\u67E5\u770B\u5F53\u524D\u8BBE\u7F6E"
+        );
+      }
+      throw error;
+    }
   }
   settings.describe = describe;
   if (originalWrite) settings.write = write;
-  return () => {
+  const dispose = () => {
     if (settings.describe === describe) settings.describe = originalDescribe;
     if (originalWrite && settings.write === write)
       settings.write = originalWrite;
   };
+  const setContextWindow = async (provider, modelId, tokens, expected) => {
+    if (!originalWrite) throw new Error("\u4E0A\u4E0B\u6587\u8BBE\u7F6E\u6682\u4E0D\u53EF\u7528");
+    if (tokens !== void 0 && !validContextWindow(tokens))
+      throw new Error("\u4E0A\u4E0B\u6587\u7A97\u53E3\u5FC5\u987B\u662F\u6B63\u7684\u5B89\u5168\u6574\u6570");
+    const edits = [];
+    const legacyRevision = contextWrites?.revision();
+    const paths = [["providers", provider, "models"]];
+    await originalWrite.call(
+      settings,
+      "llm-pi-ai",
+      (current) => {
+        const next = structuredClone(current);
+        const model = next.providers?.[provider]?.models?.find(
+          (row) => row.id === modelId
+        );
+        if (!model) throw new Error("\u6A21\u578B\u5DF2\u53D8\u66F4\uFF0C\u8BF7\u5237\u65B0\u540E\u91CD\u8BD5");
+        edits.push({ provider, modelId, previousTokens: model.contextWindow });
+        if (tokens === void 0) delete model.contextWindow;
+        else model.contextWindow = tokens;
+        contextWrites?.validate?.(provider, model, next.providers[provider]);
+        return next;
+      },
+      expected,
+      paths
+    );
+    await finishContextWrite.call(
+      settings,
+      "llm-pi-ai",
+      edits,
+      legacyRevision,
+      paths
+    );
+  };
+  return Object.assign(dispose, { setContextWindow });
 }
 
 // src/codex.ts
@@ -2299,14 +2524,15 @@ var Config2 = z2.object({
     mcpAllowlist: z2.array(String).default([])
   }).default({})
 });
-var _speed_dec, _setSpeed_dec, _removeStale_dec, _refresh_dec, _updateOptions_dec, _configure_dec, _status_dec, _refreshCodex_dec, _codexStatus_dec, _a, _init;
-var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refreshCodex_dec = [Remote], _status_dec = [Remote], _configure_dec = [Remote], _updateOptions_dec = [Remote], _refresh_dec = [Remote], _removeStale_dec = [Remote], _setSpeed_dec = [Remote({ resources: { session: (sessionId) => sessionId } })], _speed_dec = [Remote({ resources: { session: (sessionId) => sessionId } })], _a) {
+var _speed_dec, _setSpeed_dec, _setContextWindow_dec, _removeStale_dec, _refresh_dec, _updateOptions_dec, _configure_dec, _status_dec, _refreshCodex_dec, _codexStatus_dec, _a, _init;
+var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refreshCodex_dec = [Remote], _status_dec = [Remote], _configure_dec = [Remote], _updateOptions_dec = [Remote], _refresh_dec = [Remote], _removeStale_dec = [Remote], _setContextWindow_dec = [Remote], _setSpeed_dec = [Remote({ resources: { session: (sessionId) => sessionId } })], _speed_dec = [Remote({ resources: { session: (sessionId) => sessionId } })], _a) {
   constructor(ctx, options, settingsNs) {
     super(ctx, "gptModels");
     this.options = options;
     this.settingsNs = settingsNs;
     __runInitializers(_init, 5, this);
     __publicField(this, "codex");
+    __publicField(this, "contextWriter");
     __publicField(this, "typertRemote");
     __publicField(this, "runners", /* @__PURE__ */ new Map());
     __publicField(this, "snapshots", /* @__PURE__ */ new Map());
@@ -2434,11 +2660,11 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
   }
   async publish(provider, snapshot) {
     if (this.closed) return;
-    const { descriptor, profile } = this.profile(provider);
+    const { profile } = this.profile(provider);
     const existing = profile.models ?? [];
     const ids = new Set(existing.map((model) => model.id));
     const added = snapshot.models.filter((model) => !model.stale && !ids.has(model.id)).map((model) => ({ id: model.id, name: model.displayName ?? model.id }));
-    if (added.length) {
+    if (added.length && !Object.keys(profile.modelOverrides ?? {}).length) {
       await this.ctx.settings.mutate(
         "llm-pi-ai",
         [
@@ -2446,15 +2672,11 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
             op: "set",
             path: ["providers", provider, "models"],
             value: [...existing, ...added].map(
-              (model) => projectModel(
-                provider,
-                model,
-                (route, row) => this.capability(route, row.id, row.name)
-              )
+              (model) => this.editableModel(provider, model, ids.has(model.id))
             )
           }
         ],
-        descriptor.revision
+        this.modelsRevision()
       );
     }
     if (this.closed) return;
@@ -2490,6 +2712,15 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
         stale: model.stale ?? false
       })) ?? [],
       capabilities,
+      contextWindows: Object.fromEntries(
+        this.routeOptions(provider)?.enabled ? (this.profile(provider).profile.models ?? []).filter(
+          (model) => this.capability(provider, model.id) !== void 0
+        ).map((model) => [
+          model.id,
+          this.contextWindow(provider, model.id)
+        ]) : []
+      ),
+      contextRevision: this.modelsRevision(),
       options: this.routeOptions(provider) ?? {
         enabled: false,
         refreshIntervalMs: 18e5,
@@ -2530,6 +2761,10 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
     const overrides = JSON.parse(overridesJSON);
     if (!overrides || typeof overrides !== "object" || Array.isArray(overrides))
       throw new Error("\u6A21\u578B\u8986\u76D6\u9700\u8981\u662F JSON \u5BF9\u8C61");
+    for (const entry of Object.values(overrides)) {
+      if (entry && typeof entry === "object" && "contextWindow" in entry && !validContextWindow(entry.contextWindow))
+        throw new Error("\u4E0A\u4E0B\u6587\u7A97\u53E3\u5FC5\u987B\u662F\u6B63\u7684\u5B89\u5168\u6574\u6570");
+    }
     const descriptor = this.ctx.settings.describe().find((entry) => entry.ns === this.settingsNs);
     if (!descriptor) throw new Error("GPT plugin settings unavailable");
     await this.ctx.settings.mutate(
@@ -2559,7 +2794,7 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
     const runner = this.runners.get(provider);
     if (!runner) throw new Error("Enable GPT adaptation first");
     const staleIds = runner.current()?.models.filter((model) => model.stale).map((model) => model.id) ?? [];
-    const { descriptor, profile } = this.profile(provider);
+    const { profile } = this.profile(provider);
     const nextModels = profile.models?.filter(
       (model) => !staleIds.includes(model.id)
     );
@@ -2571,18 +2806,128 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
             op: "set",
             path: ["providers", provider, "models"],
             value: nextModels.map(
-              (model) => projectModel(
-                provider,
-                model,
-                (route, row) => this.capability(route, row.id, row.name)
-              )
+              (model) => this.editableModel(provider, model)
             )
           }
         ],
-        descriptor.revision
+        this.modelsRevision()
       );
     await runner.removeStale(staleIds);
   }
+  /** Internal model-list writes must use the same manual fields as native drafts. */
+  editableModel(provider, model, existing = true) {
+    const capability = this.capability(provider, model.id, model.name);
+    if (!capability) return model;
+    const tokens = existing ? this.editableContextWindow(provider, model) : model.contextWindow;
+    return projectModel(
+      provider,
+      tokens === void 0 ? model : { ...model, contextWindow: tokens },
+      () => capability,
+      void 0,
+      false
+    );
+  }
+  editableContextWindow(provider, model) {
+    const context = this.contextWindow(provider, model.id);
+    return context.source === "legacy-override" ? context.effectiveTokens : model.contextWindow;
+  }
+  /** Resolve raw metadata separately from merged capabilities and cached overrides. */
+  contextWindow(provider, modelId, nativeDefault, profileOverride) {
+    const { descriptor, profile: storedProfile } = this.profile(provider);
+    const inherited = descriptor.base?.providers?.[provider];
+    const profile = profileOverride ? { ...inherited, ...profileOverride } : storedProfile;
+    const overrides = JSON.parse(
+      this.routeOptions(provider, profile)?.overridesJSON ?? "{}"
+    );
+    const row = profile.models?.length ? profile.models.find((model) => model.id === modelId) : profile.modelOverrides?.[modelId];
+    const snapshot = this.snapshots.get(provider);
+    const raw = snapshot?.models.find((model) => model.id === modelId);
+    const userProfile = profileOverride ?? descriptor.user?.providers?.[provider];
+    const providerDefault = userProfile?.defaultContextWindow ?? (profile.defaultContextWindow !== 1e6 || descriptor.base === void 0 && descriptor.user === void 0 ? profile.defaultContextWindow : void 0);
+    const catalogModel = getBuiltinModels(provider).find(
+      (model) => model.id === modelId
+    );
+    return resolveContextWindow({
+      modelId,
+      legacy: overrides[modelId]?.contextWindow,
+      native: row?.contextWindow,
+      providerDefault,
+      upstream: raw?.capabilities?.contextWindow,
+      catalog: snapshot?.catalog?.models[modelId]?.contextWindow,
+      nativeDefault: nativeDefault ?? catalogModel?.contextWindow ?? profile.defaultContextWindow
+    });
+  }
+  /** Both native form saves and the compatibility RPC validate the same budget. */
+  validateContextWindow(provider, model, profile) {
+    const effective = model.contextWindow ?? this.contextWindow(provider, model.id, void 0, profile).defaultTokens;
+    const compaction = this.nativeDescribe().find(
+      (entry) => entry.ns === "compaction-basic"
+    )?.value;
+    const policy = compaction?.modelPolicies?.find(
+      (entry) => entry.provider === provider && entry.model === model.id
+    );
+    const headroom = policy?.headroomTokens ?? compaction?.headroomTokens ?? 0;
+    const inherited = this.nativeDescribe().find((entry) => entry.ns === "llm-pi-ai")?.base?.providers?.[provider];
+    const defaults = NativeConfig({
+      providers: {
+        [provider]: {
+          defaultMaxTokens: profile.defaultMaxTokens ?? inherited?.defaultMaxTokens
+        }
+      }
+    }).providers.get()[provider];
+    const output = model.maxTokens ?? getBuiltinModels(provider).find((entry) => entry.id === model.id)?.maxTokens ?? defaults.defaultMaxTokens;
+    if (effective <= output + headroom)
+      throw new Error("\u4E0A\u4E0B\u6587\u7A97\u53E3\u5FC5\u987B\u5927\u4E8E\u5F53\u524D\u6700\u5927\u8F93\u51FA\u548C\u538B\u7F29\u9884\u7559\u7684\u603B\u548C");
+  }
+  contextWriteRevision() {
+    return this.nativeDescribe().find((row) => row.ns === this.settingsNs)?.revision;
+  }
+  /** Public form revisions also cover changes to projected model capacities. */
+  modelsRevision() {
+    const descriptor = this.ctx.settings.describe().find((row) => row.ns === "llm-pi-ai");
+    if (!descriptor) throw new Error("\u6A21\u578B\u8BBE\u7F6E\u6682\u4E0D\u53EF\u7528");
+    return descriptor.revision;
+  }
+  /** Remove only legacy context fields after the matching native save commits. */
+  async clearLegacyContext(edits, expected) {
+    const operations = [];
+    for (const provider of new Set(edits.map((edit) => edit.provider))) {
+      const text = this.options.routes[provider]?.overridesJSON;
+      if (!text) continue;
+      const overrides = JSON.parse(text);
+      let changed = false;
+      for (const edit of edits.filter((entry) => entry.provider === provider)) {
+        if (overrides[edit.modelId] && Object.hasOwn(overrides[edit.modelId], "contextWindow")) {
+          delete overrides[edit.modelId].contextWindow;
+          changed = true;
+        }
+      }
+      if (changed)
+        operations.push({
+          op: "set",
+          path: ["routes", provider, "overridesJSON"],
+          value: JSON.stringify(overrides)
+        });
+    }
+    if (operations.length)
+      await this.ctx.settings.mutate(this.settingsNs, operations, expected);
+  }
+  async setContextWindow(provider, modelId, mode, customTokens, expectedRevision) {
+    const tokens = contextWindowTokens(mode, customTokens);
+    const { profile } = this.profile(provider);
+    if (!this.capability(provider, modelId))
+      throw new Error("\u8BF7\u5148\u542F\u7528\u8BE5 GPT \u6A21\u578B\u7684\u81EA\u52A8\u9002\u914D");
+    const index = profile.models?.findIndex((model) => model.id === modelId) ?? -1;
+    if (index < 0) throw new Error("\u6A21\u578B\u5DF2\u53D8\u66F4\uFF0C\u8BF7\u5237\u65B0\u540E\u91CD\u8BD5");
+    if (!this.contextWriter) throw new Error("\u4E0A\u4E0B\u6587\u8BBE\u7F6E\u6682\u4E0D\u53EF\u7528");
+    await this.contextWriter(
+      provider,
+      modelId,
+      tokens,
+      expectedRevision ?? this.modelsRevision()
+    );
+    this.ctx.emit("llm/models-updated", provider);
+  }
   async setSpeed(sessionId, provider, modelId, tier) {
     const capability = this.capability(provider, modelId);
     if (!capability?.serviceTiers.some((choice) => choice.id === tier))
@@ -2619,11 +2964,15 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
       ...overrides[modelId]
     };
     const current = this.snapshots.get(provider)?.capabilities[modelId];
-    return current ? mergeCapabilities(current, override, current.capabilityRevision) : resolveCapabilities(
+    const result = current ? mergeCapabilities(current, override, current.capabilityRevision) : resolveCapabilities(
       { id: modelId, ...displayName ? { displayName } : {} },
       "gpt-preset",
       override
     );
+    return {
+      ...result,
+      contextWindow: this.contextWindow(provider, modelId).effectiveTokens
+    };
   }
   /** Capture request controls for the version-checked native runtime extension. */
   capture(provider, modelId, base) {
@@ -2636,6 +2985,14 @@ var GptModels = class extends (_a = Service, _codexStatus_dec = [Remote], _refre
     const protocol = profile.api ?? "openai-responses";
     const model = {
       ...base,
+      context: {
+        ...base.context ?? {},
+        contextWindow: this.contextWindow(
+          provider,
+          modelId,
+          base.context?.contextWindow
+        ).effectiveTokens
+      },
       name: capabilities.displayName === modelId ? base.name : capabilities.displayName
     };
     if (!preserveReasoning) delete model.reasoning;
@@ -2686,6 +3043,7 @@ __decorateElement(_init, 1, "configure", _configure_dec, GptModels);
 __decorateElement(_init, 1, "updateOptions", _updateOptions_dec, GptModels);
 __decorateElement(_init, 1, "refresh", _refresh_dec, GptModels);
 __decorateElement(_init, 1, "removeStale", _removeStale_dec, GptModels);
+__decorateElement(_init, 1, "setContextWindow", _setContextWindow_dec, GptModels);
 __decorateElement(_init, 1, "setSpeed", _setSpeed_dec, GptModels);
 __decorateElement(_init, 1, "speed", _speed_dec, GptModels);
 __decoratorMetadata(_init, GptModels);
@@ -2725,9 +3083,17 @@ async function apply2(ctx, config) {
     const providers = service.enabledProviders();
     const dispose = installSettingsOverlay(
       ctx.settings,
-      (provider, model, profile) => typeof model.id === "string" ? service.capability(provider, model.id, model.name, profile) : void 0
+      (provider, model, profile) => typeof model.id === "string" ? service.capability(provider, model.id, model.name, profile) : void 0,
+      {
+        editableContext: (provider, model) => service.editableContextWindow(provider, model),
+        validate: (provider, model, profile) => service.validateContextWindow(provider, model, profile),
+        revision: () => service.contextWriteRevision(),
+        saved: (edits, revision) => service.clearLegacyContext(edits, revision)
+      }
     );
+    service.contextWriter = dispose.setContextWindow;
     return () => {
+      service.contextWriter = void 0;
       dispose();
       for (const provider of providers) ctx.emit("llm/models-updated", provider);
     };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-gpt-adapter",
-  "version": "0.2.3",
+  "version": "0.2.5",
   "private": false,
   "publishConfig": {
     "access": "public"