diff --git a/docs-site/public/pr-screenshots/1991-models-custom-windows.png b/docs-site/public/pr-screenshots/1991-models-custom-windows.png
new file mode 100644
index 0000000000..d771785175
Binary files /dev/null and b/docs-site/public/pr-screenshots/1991-models-custom-windows.png differ
diff --git a/gui/src/i18n/de.ts b/gui/src/i18n/de.ts
index f06b217022..231b46dfec 100644
--- a/gui/src/i18n/de.ts
+++ b/gui/src/i18n/de.ts
@@ -493,7 +493,7 @@ export const de: Record<TKey, string> = {
   "models.capApplied": "Kontext-Limit angewendet — greift bei der nächsten Codex-Runde.",
   "models.capSaveFailed": "Kontext-Limit konnte nicht gespeichert werden",
   "models.contextCapped": "350k-Limit",
-  "models.contextCapLabel": "Kontext-Limit",
+  "models.contextCapLabel": "Standardfenster / Limit",
   "models.v2Label": "Sub-Agent",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "Was ist v1 / v2?",
@@ -514,13 +514,13 @@ export const de: Record<TKey, string> = {
   "models.v2ThreadsApplied": "Thread-Limit aktualisiert — gilt für neue Sitzungen",
   "models.v2ThreadsInvalid": "Thread-Limit muss eine ganze Zahl >= 1 sein",
   "models.v2ThreadsApply": "Anwenden",
-  "models.capValue": "Limit {value}",
-  "models.contextSettings": "Kontextfenster",
-  "models.contextSettingsTitle": "Kontextfenster — {provider}",
+  "models.capValue": "Standard {value}",
+  "models.contextSettings": "Eigene Fenster",
+  "models.contextSettingsTitle": "Eigene Fenster — {provider}",
   "models.contextDefault": "Anbieterstandard",
   "models.contextModel": "Modell",
   "models.contextModelOverride": "Modellüberschreibung",
-  "models.contextHint": "Wird verwendet, wenn Upstream-Metadaten fehlen; andernfalls begrenzt der Wert ein größeres gemeldetes Fenster. Leer lassen für automatische Erkennung.",
+  "models.contextHint": "Wenn das Fenster bekannt ist, tragen Sie hier das tatsächliche Codex-Fenster ein. Fehlt ein Upstream-Wert, gilt dieser Eintrag; ein größeres gemeldetes Fenster wird nur nach unten begrenzt, ein kleineres bleibt erhalten. Leer bedeutet das Anbieter-«Standardfenster / Limit», oder 128k wenn das Limit aus ist.",
   "models.contextAutomatic": "Automatische Erkennung",
   "models.contextSaved": "Kontextfenster aktualisiert — gilt ab der nächsten Codex-Runde.",
   "models.contextUnchanged": "Keine Änderungen am Kontextfenster zu speichern.",
@@ -528,7 +528,7 @@ export const de: Record<TKey, string> = {
   "models.contextInvalid": "Kontextfenster müssen positive ganze Zahlen sein",
   "models.contextCappedValue": "{value}-Limit",
   "models.setAll": "Alle setzen",
-  "models.setAllHint": "Wendet das {value}-Kontext-Limit auf alle gerouteten Anbieter an. Native Anbieter bleiben unberührt.",
+  "models.setAllHint": "Schaltet das Standardfenster {value} für alle gerouteten Anbieter ein. Fehlen context_window / context_length, wird dieser Wert das tatsächliche Codex-Fenster. Für ein einzelnes Modell nutzen Sie «Eigene Fenster» in derselben Zeile. Native Anbieter bleiben unberührt.",
   "models.collapseAll": "Alle einklappen",
   "models.expandAll": "Alle ausklappen",
   "models.orderHint": "Reihenfolge in der Modellauswahl: Subagents-Auswahl (in der festgelegten Reihenfolge) → übrige geroutete Modelle alphabetisch nach Anbieter, dann Modell-ID → native Modelle. Sichtbarkeitsschalter filtern nur; sie ändern diese Reihenfolge nicht.",
diff --git a/gui/src/i18n/en.ts b/gui/src/i18n/en.ts
index d1e5389165..4bb60733fd 100644
--- a/gui/src/i18n/en.ts
+++ b/gui/src/i18n/en.ts
@@ -518,7 +518,7 @@ export const en = {
   "models.capApplied": "Context cap applied — takes effect on the next Codex turn.",
   "models.capSaveFailed": "Failed to save context cap",
   "models.contextCapped": "350k cap",
-  "models.contextCapLabel": "Context cap",
+  "models.contextCapLabel": "Default window / cap",
   "models.v2Label": "Sub-agent",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "What is v1 / v2?",
@@ -539,13 +539,13 @@ export const en = {
   "models.v2ThreadsApplied": "Thread limit updated — applies to new sessions",
   "models.v2ThreadsInvalid": "Thread limit must be an integer >= 1",
   "models.v2ThreadsApply": "Apply",
-  "models.capValue": "Cap {value}",
-  "models.contextSettings": "Context windows",
-  "models.contextSettingsTitle": "Context windows — {provider}",
+  "models.capValue": "Default {value}",
+  "models.contextSettings": "Custom windows",
+  "models.contextSettingsTitle": "Custom windows — {provider}",
   "models.contextDefault": "Provider default",
   "models.contextModel": "Model",
   "models.contextModelOverride": "Model override",
-  "models.contextHint": "Used when upstream metadata is missing; otherwise limits a larger reported window. Leave blank for automatic discovery.",
+  "models.contextHint": "Write the actual Codex window here when you already know it. This fills a missing upstream window and only lowers a larger reported one. Leave blank to use the provider Default window / cap, or 128k if that cap is off.",
   "models.contextAutomatic": "Automatic discovery",
   "models.contextSaved": "Context windows updated — takes effect on the next Codex turn.",
   "models.contextUnchanged": "No context window changes to save.",
@@ -553,7 +553,7 @@ export const en = {
   "models.contextInvalid": "Context windows must be positive whole numbers",
   "models.contextCappedValue": "{value} cap",
   "models.setAll": "Set all",
-  "models.setAllHint": "Apply the {value} context cap to every routed provider. Native providers are unaffected.",
+  "models.setAllHint": "Turn on the {value} default window for every routed provider. Relays that omit context_window / context_length get this as the actual Codex window. Use Custom windows on a provider row to set one model by hand. Native providers are unaffected.",
   "models.collapseAll": "Collapse all",
   "models.expandAll": "Expand all",
   "models.orderHint": "Picker order: Subagents picks (in the selected order) → remaining routed models alphabetically by provider, then model ID → native models. Visibility switches only filter models; they do not change this order.",
diff --git a/gui/src/i18n/fr.ts b/gui/src/i18n/fr.ts
index a6a34e5954..36aa8a0a01 100644
--- a/gui/src/i18n/fr.ts
+++ b/gui/src/i18n/fr.ts
@@ -503,7 +503,7 @@ export const fr: Record<TKey, string> = {
   "models.capApplied": "Plafond de contexte appliqué — il prendra effet au prochain tour Codex.",
   "models.capSaveFailed": "Échec de l’enregistrement du plafond de contexte",
   "models.contextCapped": "Plafond de 350k",
-  "models.contextCapLabel": "Plafond de contexte",
+  "models.contextCapLabel": "Fenêtre par défaut / plafond",
   "models.v2Label": "Sous-agent",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "Que sont v1 et v2 ?",
@@ -524,13 +524,13 @@ export const fr: Record<TKey, string> = {
   "models.v2ThreadsApplied": "Limite de fils mise à jour — s’applique aux nouvelles sessions",
   "models.v2ThreadsInvalid": "La limite de fils doit être un entier >= 1",
   "models.v2ThreadsApply": "Appliquer",
-  "models.capValue": "Plafond {value}",
-  "models.contextSettings": "Fenêtres de contexte",
-  "models.contextSettingsTitle": "Fenêtres de contexte — {provider}",
+  "models.capValue": "Défaut {value}",
+  "models.contextSettings": "Fenêtres perso",
+  "models.contextSettingsTitle": "Fenêtres perso — {provider}",
   "models.contextDefault": "Valeur par défaut du fournisseur",
   "models.contextModel": "Modèle",
   "models.contextModelOverride": "Remplacement pour le modèle",
-  "models.contextHint": "Utilisé lorsque les métadonnées en amont sont absentes ; sinon, limite une fenêtre déclarée plus grande. Laisser vide pour une détection automatique.",
+  "models.contextHint": "Si vous connaissez déjà la fenêtre, écrivez ici la fenêtre Codex réelle. Sans métadonnées amont, cette valeur est utilisée ; une fenêtre plus grande est seulement abaissée, une plus petite est conservée. Laisser vide utilise la « Fenêtre par défaut / plafond » du fournisseur, ou 128k si ce plafond est désactivé.",
   "models.contextAutomatic": "Détection automatique",
   "models.contextSaved": "Fenêtres de contexte mises à jour — prend effet au prochain tour Codex.",
   "models.contextUnchanged": "Aucune modification des fenêtres de contexte à enregistrer.",
@@ -538,7 +538,7 @@ export const fr: Record<TKey, string> = {
   "models.contextInvalid": "Les fenêtres de contexte doivent être des nombres entiers positifs",
   "models.contextCappedValue": "Plafond de {value}",
   "models.setAll": "Tout définir",
-  "models.setAllHint": "Appliquer le plafond de contexte de {value} à chaque fournisseur routé. Les fournisseurs natifs ne sont pas affectés.",
+  "models.setAllHint": "Active la fenêtre par défaut {value} pour chaque fournisseur routé. Si un relais omet context_window / context_length, cette valeur devient la fenêtre Codex réelle. Pour un seul modèle, utilisez « Fenêtres perso » sur la même ligne. Les fournisseurs natifs ne sont pas affectés.",
   "models.collapseAll": "Tout réduire",
   "models.expandAll": "Tout développer",
   "models.orderHint": "Ordre du sélecteur : choix des sous-agents (dans l’ordre sélectionné) → autres modèles routés, classés par ordre alphabétique du fournisseur puis par ID de modèle → modèles natifs. Les options de visibilité ne font que filtrer les modèles ; elles ne modifient pas cet ordre.",
diff --git a/gui/src/i18n/ja.ts b/gui/src/i18n/ja.ts
index 23889a0742..9847db888d 100644
--- a/gui/src/i18n/ja.ts
+++ b/gui/src/i18n/ja.ts
@@ -501,7 +501,7 @@ export const ja: Record<TKey, string> = {
   "models.capApplied": "コンテキスト上限を適用しました — 次回の Codex ターンで有効になります。",
   "models.capSaveFailed": "コンテキスト上限の保存に失敗しました",
   "models.contextCapped": "350k 上限",
-  "models.contextCapLabel": "コンテキスト上限",
+  "models.contextCapLabel": "デフォルトウィンドウ / 上限",
   "models.v2Label": "サブエージェント",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "v1 / v2 とは?",
@@ -522,13 +522,13 @@ export const ja: Record<TKey, string> = {
   "models.v2ThreadsApplied": "スレッド上限を更新しました — 新規セッションに適用",
   "models.v2ThreadsInvalid": "スレッド上限は 1 以上の整数にしてください",
   "models.v2ThreadsApply": "適用",
-  "models.capValue": "上限 {value}",
-  "models.contextSettings": "コンテキストウィンドウ",
-  "models.contextSettingsTitle": "コンテキストウィンドウ — {provider}",
+  "models.capValue": "デフォルト {value}",
+  "models.contextSettings": "カスタムウィンドウ",
+  "models.contextSettingsTitle": "カスタムウィンドウ — {provider}",
   "models.contextDefault": "プロバイダーのデフォルト",
   "models.contextModel": "モデル",
   "models.contextModelOverride": "モデル別の上書き",
-  "models.contextHint": "上流メタデータがない場合に使われ、メタデータがある場合は報告値の上限になります。自動検出に戻すには空欄にします。",
+  "models.contextHint": "すでに分かっている場合は、ここに Codex の実際のコンテキストウィンドウを書きます。上流の値がないときはこの値を使い、より大きな報告値だけ下げ、より小さな上流のコンテキストウィンドウはそのまま残します。空欄ならプロバイダーの「デフォルトウィンドウ / 上限」を使い、その上限がオフなら 128k です。",
   "models.contextAutomatic": "自動検出",
   "models.contextSaved": "コンテキストウィンドウを更新しました — 次回の Codex ターンから有効です。",
   "models.contextUnchanged": "保存するコンテキストウィンドウの変更はありません。",
@@ -536,7 +536,7 @@ export const ja: Record<TKey, string> = {
   "models.contextInvalid": "コンテキストウィンドウは正の整数で指定してください",
   "models.contextCappedValue": "{value} 上限",
   "models.setAll": "すべて設定",
-  "models.setAllHint": "{value} のコンテキスト上限をすべてのルーティング済みプロバイダーに適用します。ネイティブプロバイダーには影響しません。",
+  "models.setAllHint": "すべてのルーティング済みプロバイダーに {value} のデフォルトウィンドウをオンにします。中継が context_window / context_length を返さない場合、この値が実際の Codex ウィンドウになります。1 モデルだけ手で書くときは同じ行の「カスタムウィンドウ」を使います。ネイティブプロバイダーには影響しません。",
   "models.collapseAll": "すべて折りたたむ",
   "models.expandAll": "すべて展開",
   "models.orderHint": "ピッカーの順序: サブエージェントの選択(選択順) → 残りのルーティングモデルはプロバイダー別、次にモデル ID 別のアルファベット順 → ネイティブモデル。表示切り替えはモデルをフィルタするだけで、この順序は変更しません。",
diff --git a/gui/src/i18n/ko.ts b/gui/src/i18n/ko.ts
index dc7fb87462..7ca9059243 100644
--- a/gui/src/i18n/ko.ts
+++ b/gui/src/i18n/ko.ts
@@ -504,7 +504,7 @@ export const ko: Record<TKey, string> = {
   "models.capApplied": "컨텍스트 제한 적용됨 — 다음 Codex 턴부터 반영됩니다.",
   "models.capSaveFailed": "컨텍스트 제한 저장 실패",
   "models.contextCapped": "350k 제한",
-  "models.contextCapLabel": "컨텍스트 제한",
+  "models.contextCapLabel": "기본 창 / 상한",
   "models.v2Label": "서브에이전트",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2Mode_v1": "v1",
@@ -525,13 +525,13 @@ export const ko: Record<TKey, string> = {
   "models.v2ThreadsApplied": "스레드 한도 변경됨 — 새 세션부터 적용",
   "models.v2ThreadsInvalid": "스레드 한도는 1 이상 정수여야 합니다",
   "models.v2ThreadsApply": "적용",
-  "models.capValue": "{value} 제한",
-  "models.contextSettings": "컨텍스트 윈도우",
-  "models.contextSettingsTitle": "컨텍스트 윈도우 — {provider}",
+  "models.capValue": "기본 {value}",
+  "models.contextSettings": "사용자 지정 창",
+  "models.contextSettingsTitle": "사용자 지정 창 — {provider}",
   "models.contextDefault": "프로바이더 기본값",
   "models.contextModel": "모델",
   "models.contextModelOverride": "모델별 재정의",
-  "models.contextHint": "업스트림 메타데이터가 없을 때 사용하며, 메타데이터가 있으면 더 큰 보고값의 상한으로 적용합니다. 자동 검색을 사용하려면 비워 두세요.",
+  "models.contextHint": "이미 아는 경우 여기에 실제 Codex 컨텍스트 윈도우를 적습니다. 업스트림 값이 없으면 이 값을 쓰고, 더 큰 보고값만 낮추며, 더 작은 업스트림 컨텍스트 윈도우는 그대로 둡니다. 비우면 프로바이더의 「기본 창 / 상한」을 쓰고, 그 상한이 꺼져 있으면 128k입니다.",
   "models.contextAutomatic": "자동 검색",
   "models.contextSaved": "컨텍스트 윈도우가 업데이트되었습니다 — 다음 Codex 턴부터 적용됩니다.",
   "models.contextUnchanged": "저장할 컨텍스트 윈도우 변경이 없습니다.",
@@ -539,7 +539,7 @@ export const ko: Record<TKey, string> = {
   "models.contextInvalid": "컨텍스트 윈도우는 양의 정수여야 합니다",
   "models.contextCappedValue": "{value} 제한",
   "models.setAll": "전체 적용",
-  "models.setAllHint": "{value} 컨텍스트 상한을 라우팅된 모든 프로바이더에 적용합니다. 네이티브 프로바이더는 영향을 받지 않습니다.",
+  "models.setAllHint": "라우팅된 모든 프로바이더에 {value} 기본 창을 켭니다. 중계가 context_window / context_length 를 주지 않으면 이 값이 실제 Codex 창이 됩니다. 모델 하나만 손으로 쓰려면 같은 줄의 「사용자 지정 창」을 쓰세요. 네이티브 프로바이더는 영향을 받지 않습니다.",
   "models.collapseAll": "모두 접기",
   "models.expandAll": "모두 펼치기",
   "models.orderHint": "피커 순서: Subagents에서 지정한 순서 → 나머지 라우팅 모델(프로바이더, 모델 ID 순 알파벳 정렬) → 네이티브 모델. 노출 토글은 모델을 필터링할 뿐 이 순서를 바꾸지 않습니다.",
diff --git a/gui/src/i18n/ru.ts b/gui/src/i18n/ru.ts
index 2368ee0257..21c33c3559 100644
--- a/gui/src/i18n/ru.ts
+++ b/gui/src/i18n/ru.ts
@@ -506,7 +506,7 @@ export const ru: Record<TKey, string> = {
   "models.capApplied": "Лимит контекста применён — вступит в силу на следующем ходе Codex.",
   "models.capSaveFailed": "Не удалось сохранить лимит контекста",
   "models.contextCapped": "Лимит 350k",
-  "models.contextCapLabel": "Лимит контекста",
+  "models.contextCapLabel": "Окно по умолчанию / лимит",
   "models.v2Label": "Подагент",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "Что такое v1 / v2?",
@@ -527,13 +527,13 @@ export const ru: Record<TKey, string> = {
   "models.v2ThreadsApplied": "Лимит потоков обновлён — применяется к новым сессиям",
   "models.v2ThreadsInvalid": "Лимит потоков должен быть целым числом >= 1",
   "models.v2ThreadsApply": "Применить",
-  "models.capValue": "Лимит {value}",
-  "models.contextSettings": "Контекстные окна",
-  "models.contextSettingsTitle": "Контекстные окна — {provider}",
+  "models.capValue": "По умолч. {value}",
+  "models.contextSettings": "Пользовательские окна",
+  "models.contextSettingsTitle": "Пользовательские окна — {provider}",
   "models.contextDefault": "Значение провайдера",
   "models.contextModel": "Модель",
   "models.contextModelOverride": "Переопределение модели",
-  "models.contextHint": "Используется, если вышестоящие метаданные отсутствуют; иначе ограничивает большее заявленное окно. Оставьте поле пустым для автоматического определения.",
+  "models.contextHint": "Если окно уже известно, запишите здесь реальное окно Codex. При отсутствии метаданных используется это значение; большее заявленное окно только ограничивается, меньшее сохраняется. Пустое поле берёт «Окно по умолчанию / лимит» провайдера или 128k, если лимит выключен.",
   "models.contextAutomatic": "Автоматическое определение",
   "models.contextSaved": "Контекстные окна обновлены — изменения вступят в силу на следующем ходе Codex.",
   "models.contextUnchanged": "Нет изменений контекстных окон для сохранения.",
@@ -541,7 +541,7 @@ export const ru: Record<TKey, string> = {
   "models.contextInvalid": "Контекстные окна должны быть положительными целыми числами",
   "models.contextCappedValue": "Лимит {value}",
   "models.setAll": "Применить ко всем",
-  "models.setAllHint": "Применяет лимит контекста {value} ко всем маршрутизируемым провайдерам. Нативные провайдеры не затрагиваются.",
+  "models.setAllHint": "Включает окно по умолчанию {value} для всех маршрутизируемых провайдеров. Если релей не отдаёт context_window / context_length, это значение становится реальным окном Codex. Чтобы задать одну модель вручную, используйте «Пользовательские окна» в той же строке. Нативные провайдеры не затрагиваются.",
   "models.collapseAll": "Свернуть все",
   "models.expandAll": "Развернуть все",
   "models.orderHint": "Порядок в селекторе: модели, выбранные на странице «Подагенты» (в заданном порядке) → остальные маршрутизируемые модели по алфавиту — сначала по провайдеру, затем по ID модели → нативные модели. Переключатели видимости лишь фильтруют модели и не меняют этот порядок.",
diff --git a/gui/src/i18n/tr.ts b/gui/src/i18n/tr.ts
index 7644e50c89..abc47076d3 100644
--- a/gui/src/i18n/tr.ts
+++ b/gui/src/i18n/tr.ts
@@ -509,7 +509,7 @@ export const tr: Record<TKey, string> = {
   "models.capApplied": "Bağlam sınırı uygulandı.",
   "models.capSaveFailed": "Bağlam sınırı kaydedilemedi",
   "models.contextCapped": "350k sınırı",
-  "models.contextCapLabel": "Bağlam sınırı",
+  "models.contextCapLabel": "Varsayılan pencere / sınır",
   "models.v2Label": "Alt Ajan",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "v1 / v2 nedir?",
@@ -530,13 +530,13 @@ export const tr: Record<TKey, string> = {
   "models.v2ThreadsApplied": "İş parçacığı limiti güncellendi",
   "models.v2ThreadsInvalid": "İş parçacığı limiti >= 1 tamsayı olmalıdır",
   "models.v2ThreadsApply": "Uygula",
-  "models.capValue": "Sınır {value}",
-  "models.contextSettings": "Bağlam pencereleri",
-  "models.contextSettingsTitle": "Bağlam pencereleri — {provider}",
+  "models.capValue": "Varsayılan {value}",
+  "models.contextSettings": "Özel pencereler",
+  "models.contextSettingsTitle": "Özel pencereler — {provider}",
   "models.contextDefault": "Sağlayıcı varsayılanı",
   "models.contextModel": "Model",
   "models.contextModelOverride": "Model geçersiz kılma",
-  "models.contextHint": "Yukarı akış meta verileri eksik olduğunda kullanılır.",
+  "models.contextHint": "Pencereyi biliyorsanız gerçek Codex penceresini buraya yazın. Üst akış değer yoksa bu kullanılır; daha büyük bildirilen pencere düşürülür, daha küçük olan korunur. Boş bırakırsanız sağlayıcının «Varsayılan pencere / sınır» değeri kullanılır; o sınır kapalıysa 128k olur.",
   "models.contextAutomatic": "Otomatik keşif",
   "models.contextSaved": "Bağlam pencereleri güncellendi.",
   "models.contextUnchanged": "Kaydedilecek bağlam penceresi değişikliği yok.",
@@ -544,7 +544,7 @@ export const tr: Record<TKey, string> = {
   "models.contextInvalid": "Bağlam pencereleri pozitif tam sayılar olmalıdır",
   "models.contextCappedValue": "{value} sınırı",
   "models.setAll": "Tümünü ayarla",
-  "models.setAllHint": "{value} bağlam sınırını her yönlendirilen sağlayıcıya uygulayın.",
+  "models.setAllHint": "Her yönlendirilen sağlayıcıda {value} varsayılan pencereyi açar. Röle context_window / context_length vermezse bu değer gerçek Codex penceresi olur. Tek bir modeli elle yazmak için aynı satırdaki «Özel pencereler»i kullanın.",
   "models.collapseAll": "Tümünü daralt",
   "models.expandAll": "Tümünü genişlet",
   "models.orderHint": "Seçici sırası: Alt ajan seçimleri → kalan modeller.",
diff --git a/gui/src/i18n/zh-TW.ts b/gui/src/i18n/zh-TW.ts
index 4f5cb8331d..9a07183f59 100644
--- a/gui/src/i18n/zh-TW.ts
+++ b/gui/src/i18n/zh-TW.ts
@@ -385,7 +385,7 @@ export const zhTW: Record<TKey, string> = {
   "models.capApplied": "上下文限制已套用 — 將在下一個 Codex 回合生效。",
   "models.capSaveFailed": "儲存上下文限制失敗",
   "models.contextCapped": "350k 限制",
-  "models.contextCapLabel": "上下文限制",
+  "models.contextCapLabel": "預設視窗 / 上限",
   "models.v2Label": "子代理",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2DocsLink": "v1 / v2 是什麼？",
@@ -406,10 +406,10 @@ export const zhTW: Record<TKey, string> = {
   "models.v2ThreadsApplied": "執行緒上限已更新 — 新會話生效",
   "models.v2ThreadsInvalid": "執行緒上限必須為 >= 1 的整數",
   "models.v2ThreadsApply": "套用",
-  "models.capValue": "限制 {value}",
+  "models.capValue": "預設 {value}",
   "models.contextCappedValue": "{value} 限制",
   "models.setAll": "全部設定",
-  "models.setAllHint": "將 {value} 上下文上限套用到所有已路由的供應商。原生供應商不受影響。",
+  "models.setAllHint": "為所有已路由供應商打開 {value} 預設視窗。中繼站沒回報 context_window / context_length 時，這個值就是 Codex 實際視窗。要幫單一模型手寫，用同一列上的「自訂視窗」。原生供應商不受影響。",
   "models.collapseAll": "全部摺疊",
   "models.expandAll": "全部展開",
   "models.orderHint": "選擇器順序：Subagents 中的選擇（按所選順序）→ 其餘已路由模型（依次按供應商、模型 ID 字母排序）→ 原生模型。可見性開關僅用於篩選，不會改變此順序。",
@@ -1771,12 +1771,12 @@ export const zhTW: Record<TKey, string> = {
   "models.subtitle.combos": "將多個模型合成一個 id 來回答。容錯移轉會依序嘗試目標；輪詢則分攤負載。",
   "models.subtitle.compatibility": "來自實驗室投影證據的唯讀相容性判定矩陣。",
   "models.subtitle.routing": "原則設定檔、dry-run 評估，以及有來源依據的路由分析。",
-  "models.contextSettings": "上下文視窗",
-  "models.contextSettingsTitle": "上下文視窗 — {provider}",
+  "models.contextSettings": "自訂視窗",
+  "models.contextSettingsTitle": "自訂視窗 — {provider}",
   "models.contextDefault": "供應商預設",
   "models.contextModel": "模型",
   "models.contextModelOverride": "模型覆寫",
-  "models.contextHint": "上游缺少中繼資料時使用此值；否則會限制較大的回報視窗。留空以自動偵測。",
+  "models.contextHint": "已經知道視窗時，在這裡手寫 Codex 實際視窗。上游沒回報視窗就用這個值；上游回報更大視窗才壓低。留空則使用供應商的「預設視窗 / 上限」；那個開關沒開時才回退 128k。",
   "models.contextAutomatic": "自動偵測",
   "models.contextSaved": "上下文視窗已更新 — 將在下一個 Codex 回合生效。",
   "models.contextUnchanged": "沒有需要儲存的上下文視窗變更。",
diff --git a/gui/src/i18n/zh.ts b/gui/src/i18n/zh.ts
index 3589644f4a..f8f147ef8d 100644
--- a/gui/src/i18n/zh.ts
+++ b/gui/src/i18n/zh.ts
@@ -501,7 +501,7 @@ export const zh: Record<TKey, string> = {
   "models.capApplied": "上下文限制已应用 — 将在下一个 Codex 回合生效。",
   "models.capSaveFailed": "保存上下文限制失败",
   "models.contextCapped": "350k 限制",
-  "models.contextCapLabel": "上下文限制",
+  "models.contextCapLabel": "默认窗口 / 上限",
   "models.v2Label": "子代理",
   "models.shadowCallOriginal": "⚠ {models} →",
   "models.v2Mode_v1": "v1",
@@ -522,13 +522,13 @@ export const zh: Record<TKey, string> = {
   "models.v2ThreadsApplied": "线程上限已更新 — 新会话生效",
   "models.v2ThreadsInvalid": "线程上限必须为 >= 1 的整数",
   "models.v2ThreadsApply": "应用",
-  "models.capValue": "限制 {value}",
-  "models.contextSettings": "上下文窗口",
-  "models.contextSettingsTitle": "上下文窗口 — {provider}",
+  "models.capValue": "默认 {value}",
+  "models.contextSettings": "自定义窗口",
+  "models.contextSettingsTitle": "自定义窗口 — {provider}",
   "models.contextDefault": "提供方默认值",
   "models.contextModel": "模型",
   "models.contextModelOverride": "模型覆盖值",
-  "models.contextHint": "上游缺少元数据时使用该值；上游已有元数据时，它只限制更大的报告值。留空则恢复自动发现。",
+  "models.contextHint": "已经知道窗口时，在这里手写 Codex 实际窗口。上游没报窗口就用这个值；上游报了更大窗口才压低。留空则使用提供方的「默认窗口 / 上限」；那个开关没开时才回退 128k。",
   "models.contextAutomatic": "自动发现",
   "models.contextSaved": "上下文窗口已更新 — 将在下一个 Codex 回合生效。",
   "models.contextUnchanged": "没有需要保存的上下文窗口更改。",
@@ -536,7 +536,7 @@ export const zh: Record<TKey, string> = {
   "models.contextInvalid": "上下文窗口必须为正整数",
   "models.contextCappedValue": "{value} 限制",
   "models.setAll": "全部设置",
-  "models.setAllHint": "将 {value} 上下文上限应用到所有已路由的提供方。原生提供方不受影响。",
+  "models.setAllHint": "给所有已路由提供方打开 {value} 默认窗口。中转站没报 context_window / context_length 时，这个值就是 Codex 实际窗口。要给单个模型手写，用同一行上的「自定义窗口」。原生提供方不受影响。",
   "models.collapseAll": "全部折叠",
   "models.expandAll": "全部展开",
   "models.orderHint": "选择器顺序：Subagents 中的选择（按所选顺序）→ 其余已路由模型（依次按提供方、模型 ID 字母排序）→ 原生模型。可见性开关仅用于筛选，不会改变此顺序。",
diff --git a/src/codex/catalog/effort.ts b/src/codex/catalog/effort.ts
index a2c462d046..3bf5daa086 100644
--- a/src/codex/catalog/effort.ts
+++ b/src/codex/catalog/effort.ts
@@ -12,7 +12,7 @@ import { CODEX_REASONING_LEVELS, codexEffortRank, configuredReasoningEfforts, mo
 import { getModelMetadata, getModelMetadataCaseInsensitive, listModelMetadata, resolveMetadataProvider } from "../../generated/model-metadata";
 import { enrichProviderFromRegistry, shouldCaseFoldMetadataModelId } from "../../providers/derive";
 import { getProviderRegistryEntry } from "../../providers/registry";
-import { applyProviderContextCap, providerContextCap } from "../../providers/context-cap";
+import { applyProviderContextCap, providerContextCap, resolveUnknownRoutedContextWindow } from "../../providers/context-cap";
 import { routedSlug, slugEquals, slugsEquivalent } from "../../providers/slug-codec";
 import { CODEX_GPT5_IDENTITY_LINE } from "../../adapters/identity";
 import { filterCursorConfiguredModelsByLiveDiscovery } from "../../adapters/cursor/discovery";
@@ -122,11 +122,14 @@ export function applyCatalogModelMetadata(entry: RawEntry, model?: CatalogModel)
   // so genuine upstream marketing names are preserved untouched.
   const displayName = typeof model.displayName === "string" ? model.displayName.trim() : "";
   if (displayName) entry.display_name = displayName;
-  if (typeof model.contextWindow === "number" && model.contextWindow > 0) {
-    entry.context_window = model.contextWindow;
-    entry.max_context_window = model.contextWindow;
+  const resolvedContext = typeof model.contextWindow === "number" && model.contextWindow > 0
+    ? model.contextWindow
+    : (model.contextCap !== undefined ? resolveUnknownRoutedContextWindow(model.contextCap) : undefined);
+  if (typeof resolvedContext === "number" && resolvedContext > 0) {
+    entry.context_window = resolvedContext;
+    entry.max_context_window = resolvedContext;
     entry.auto_compact_token_limit = Math.min(
-      Math.floor(model.contextWindow * 0.9),
+      Math.floor(resolvedContext * 0.9),
       model.maxInputTokens ?? Number.POSITIVE_INFINITY,
     );
   }
diff --git a/src/codex/catalog/provider-fetch.ts b/src/codex/catalog/provider-fetch.ts
index 2675278d07..f1acb1f699 100644
--- a/src/codex/catalog/provider-fetch.ts
+++ b/src/codex/catalog/provider-fetch.ts
@@ -39,7 +39,7 @@ import {
 } from "../../providers/service-tier";
 import { effectiveGoogleMode, getProviderRegistryEntry, providerMatchesRegistryTransport } from "../../providers/registry";
 import { parseAntigravityAvailableModels, registerAntigravityDiscoveredWireModels } from "../../providers/antigravity-models";
-import { applyProviderContextCap, providerContextCap } from "../../providers/context-cap";
+import { applyProviderContextCap, providerContextCap, resolveUnknownRoutedContextWindow } from "../../providers/context-cap";
 import { routedSlug, slugEquals, slugsEquivalent } from "../../providers/slug-codec";
 import { CODEX_GPT5_IDENTITY_LINE } from "../../adapters/identity";
 import { filterCursorConfiguredModelsByLiveDiscovery } from "../../adapters/cursor/discovery";
@@ -648,15 +648,16 @@ export function applyProviderConfigHints(name: string, prov: OcxProviderConfig,
   const supportsReasoningSummaries = configuredReasoningSummarySupport(prov, model.id);
   const supportsServiceTier = serviceTierSupportForModel(prov, model.id, name);
   const { supportsServiceTier: _staleServiceTier, ...modelWithoutServiceTier } = model;
+  // 已发现窗口只允许被配置值压低；缺窗口时，已开的 Context cap 就是实际窗口。
+  const discoveredWindow = typeof model.contextWindow === "number" && model.contextWindow > 0
+    ? model.contextWindow
+    : undefined;
+  const hintedWindow = discoveredWindow !== undefined
+    ? (configuredCap !== undefined ? Math.min(discoveredWindow, configuredCap) : discoveredWindow)
+    : (configuredCap ?? (providerCap !== undefined ? resolveUnknownRoutedContextWindow(providerCap) : undefined));
   const hinted = {
     ...modelWithoutServiceTier,
-    ...(configuredCap !== undefined
-      ? {
-        contextWindow: typeof model.contextWindow === "number" && model.contextWindow > 0
-          ? Math.min(model.contextWindow, configuredCap)
-          : configuredCap,
-      }
-      : {}),
+    ...(hintedWindow !== undefined ? { contextWindow: hintedWindow } : {}),
     ...(inputModalities ? { inputModalities } : {}),
     ...(reasoningEfforts !== undefined ? { reasoningEfforts } : {}),
     ...(configuredMaxInput !== undefined
@@ -695,9 +696,11 @@ export function applyConfigHintsToCachedModels(name: string, prov: OcxProviderCo
 
 
 /**
- * Last-resort context window for combo member synthesis when discovery and
- * provider config both omit one. Matches the catalog entry default in
- * `normalizeRoutedCatalogEntry` so incomplete live rows still catalog.
+ * Last-resort context window for combo member synthesis when discovery,
+ * provider config, and an enabled Context cap all omit one. Matches the
+ * catalog entry default in `normalizeRoutedCatalogEntry` so incomplete live
+ * rows still catalog. An enabled Context cap is the operator-facing window,
+ * not a clamp on this placeholder.
  */
 const COMBO_MEMBER_CONTEXT_FALLBACK = 128_000;
 
@@ -715,9 +718,9 @@ interface ComboCatalogMemberFallback {
  * lacks a positive contextWindow, synthesize from the (registry-enriched)
  * provider config so combos remain catalogued when targets are configured but
  * discovery metadata is incomplete. Disabled providers stay unresolved.
- * When hints still omit contextWindow, prefer known maxInputTokens, else
- * COMBO_MEMBER_CONTEXT_FALLBACK so a live row without ctx does not drop the
- * whole combo from the public catalog.
+ * When hints still omit contextWindow, prefer known maxInputTokens, else the
+ * enabled Context cap, else COMBO_MEMBER_CONTEXT_FALLBACK so a live row
+ * without ctx does not drop the whole combo from the public catalog.
  */
 export function resolveComboCatalogMember(
   target: { provider: string; model: string },
@@ -806,12 +809,15 @@ export function resolveComboCatalogMember(
   const uncappedContext = hintedContext
     ?? knownMaxInput
     ?? fallbackContext
-    ?? (existing || prov ? COMBO_MEMBER_CONTEXT_FALLBACK : undefined);
+    ?? (existing || prov ? resolveUnknownRoutedContextWindow(contextCap) : undefined);
   if (uncappedContext === undefined) return undefined;
-  const usedFallback = hintedContext === undefined;
-  const cappedContext = applyProviderContextCap(uncappedContext, contextCap);
+  // 真发现值才压低。resolveUnknownRoutedContextWindow 已经把 cap 当成窗口填进去了，不能再 min 一次。
+  const usedDiscoveredWindow = hintedContext !== undefined || knownMaxInput !== undefined || fallbackContext !== undefined;
+  const cappedContext = usedDiscoveredWindow
+    ? applyProviderContextCap(uncappedContext, contextCap)
+    : uncappedContext;
   const contextWindow = cappedContext ?? uncappedContext;
-  const fallbackCapped = usedFallback
+  const fallbackCapped = usedDiscoveredWindow
     && contextCap !== undefined
     && cappedContext !== undefined
     && cappedContext !== uncappedContext;
diff --git a/src/codex/catalog/sync.ts b/src/codex/catalog/sync.ts
index 9ee2886028..6c4b859060 100644
--- a/src/codex/catalog/sync.ts
+++ b/src/codex/catalog/sync.ts
@@ -307,8 +307,8 @@ export function deriveEntry(
     if (isRouted) {
       // A routed model is NOT the native template: never inherit its context
       // window when /models omits context metadata (#992). Known metadata
-      // restores exact values below; otherwise the strict-fields fallback
-      // supplies the conservative 128k triple.
+      // restores exact values below; an enabled Context cap fills the gap;
+      // otherwise the strict-fields fallback supplies the 128k triple.
       if (!codexForwardNativeCapabilityAlias) {
         delete e.context_window;
         delete e.max_context_window;
diff --git a/src/providers/context-cap.ts b/src/providers/context-cap.ts
index 4a4ac3f376..24834bb522 100644
--- a/src/providers/context-cap.ts
+++ b/src/providers/context-cap.ts
@@ -27,6 +27,15 @@ export function applyProviderContextCap(contextWindow: number | undefined, cap:
   return contextWindow > cap ? cap : contextWindow;
 }
 
+/**
+ * 上游没报窗口时，已开启的 Context cap 就是实际窗口。
+ * 128k 只是 Codex 解析器的兼容底线，不能当成“已发现窗口”再拿去和 cap 做 min。
+ */
+export function resolveUnknownRoutedContextWindow(cap: number | undefined): number {
+  const window = isValidContextCap(cap) ? Math.floor(cap) : 0;
+  return window > 0 ? window : 128_000;
+}
+
 /** Effective global cap value: explicit config value, else the built-in default. */
 export function globalContextCapValue(config: Pick<OcxConfig, "contextCapValue">): number {
   const value = config.contextCapValue;
diff --git a/tests/codex-catalog.test.ts b/tests/codex-catalog.test.ts
index 7086b6850a..d41daa58ad 100644
--- a/tests/codex-catalog.test.ts
+++ b/tests/codex-catalog.test.ts
@@ -1441,7 +1441,7 @@ describe("combo catalog capability intersection", () => {
       maxInputTokens: 128_000,
       inputModalities: ["text"],
     });
-    // Provider contextCap below the 128k fallback clamps the synthesized window.
+    // Provider contextCap below the 128k fallback fills the unknown window.
     expect(resolveComboCatalogMember(
       { provider: "a", model: "ghost" },
       new Map(),
@@ -1453,9 +1453,8 @@ describe("combo catalog capability intersection", () => {
       contextWindow: 64_000,
       maxInputTokens: 64_000,
       contextCap: 64_000,
-      contextCapped: true,
     });
-    // Cap above the fallback leaves 128k (no artificial raise, no capped flag).
+    // Cap above the empty discovery window fills that window. This is not a clamp.
     const aboveCap = resolveComboCatalogMember(
       { provider: "a", model: "ghost" },
       new Map(),
@@ -1463,10 +1462,10 @@ describe("combo catalog capability intersection", () => {
       200_000,
     );
     expect(aboveCap).toMatchObject({
-      contextWindow: 128_000,
-      maxInputTokens: 128_000,
+      contextWindow: 200_000,
+      maxInputTokens: 200_000,
+      contextCap: 200_000,
     });
-    // Cap may be recorded for bookkeeping (contextCapped: false) but must not claim a clamp.
     expect(aboveCap?.contextCapped).toBeFalsy();
     // No provider entry and no discovery row — cannot invent a member.
     expect(resolveComboCatalogMember(
@@ -4299,12 +4298,14 @@ describe("Codex catalog routed normalization", () => {
     expect(routed?.auto_compact_token_limit).toBe(115_200);
   });
 
-  test("a provider context cap never invents routed capacity (#992)", () => {
+  test("a provider context cap fills an unknown routed window (#992)", () => {
     const entries = buildCatalogEntries({ context_window: 372_000 }, [], [
-      { provider: "relay", id: "relay-model", contextCap: 950_000 },
+      { provider: "relay", id: "relay-model", contextCap: 350_000 },
     ]);
     const routed = entries.find(e => e.slug === "relay/relay-model");
-    expect(routed?.context_window).toBe(128_000);
+    expect(routed?.context_window).toBe(350_000);
+    expect(routed?.max_context_window).toBe(350_000);
+    expect(routed?.auto_compact_token_limit).toBe(315_000);
   });
 
   test("known routed metadata still restores the exact context window (#992)", () => {
@@ -4752,6 +4753,37 @@ describe("Codex catalog routed normalization", () => {
     expect(routed?.auto_compact_token_limit).toBe(115_200);
   });
 
+  test("an id-only model with an enabled context cap uses that cap as the window", async () => {
+    globalThis.fetch = (async () => new Response(
+      JSON.stringify({ data: [{ id: "gpt-5.6-terra" }] }),
+      { status: 200, headers: { "content-type": "application/json" } },
+    )) as typeof fetch;
+
+    const models = await gatherRoutedModels({
+      port: 10100,
+      defaultProvider: "sub2api",
+      providerContextCaps: { sub2api: 350_000 },
+      providers: {
+        sub2api: {
+          adapter: "openai-chat",
+          baseUrl: "https://sub2api.test/v1",
+          apiKey: "sk-test",
+        },
+      },
+    });
+    const routed = buildCatalogEntries(nativeTemplate(), [], models)
+      .find(e => e.slug === "sub2api/gpt-5.6-terra");
+
+    expect(models.find(m => m.id === "gpt-5.6-terra")).toMatchObject({
+      contextWindow: 350_000,
+      contextCap: 350_000,
+      contextCapped: false,
+    });
+    expect(routed?.context_window).toBe(350_000);
+    expect(routed?.max_context_window).toBe(350_000);
+    expect(routed?.auto_compact_token_limit).toBe(315_000);
+  });
+
   test("upstream metadata smaller than the configured window wins (#1073)", async () => {
     globalThis.fetch = (async () => new Response(
       JSON.stringify({ data: [{ id: "gpt-5.6-luna", context_length: 64_000 }] }),
@@ -4880,7 +4912,7 @@ describe("Codex catalog routed normalization", () => {
     });
   });
 
-  test("provider context-cap toggle lowers only known windows above 350k", async () => {
+  test("provider context-cap toggle lowers known windows above 350k and fills unknown ones", async () => {
     globalThis.fetch = (async () => new Response(JSON.stringify({
       data: [
         { id: "wide-model", metadata: { limits: { max_context_length: 500_000 } } },
@@ -4913,13 +4945,13 @@ describe("Codex catalog routed normalization", () => {
       contextCapped: false,
     });
     expect(models.find(m => m.id === "unknown-model")).toMatchObject({
+      contextWindow: 350_000,
       contextCap: 350_000,
       contextCapped: false,
     });
-    expect(models.find(m => m.id === "unknown-model")?.contextWindow).toBeUndefined();
   });
 
-  test("provider context-cap toggle does not invent context for static no-metadata models", async () => {
+  test("provider context-cap toggle fills static no-metadata models", async () => {
     let fetchCalls = 0;
     globalThis.fetch = (() => {
       fetchCalls += 1;
@@ -4943,10 +4975,10 @@ describe("Codex catalog routed normalization", () => {
 
     expect(fetchCalls).toBe(0);
     expect(models.find(m => m.id === "static-no-context")).toMatchObject({
+      contextWindow: 350_000,
       contextCap: 350_000,
       contextCapped: false,
     });
-    expect(models.find(m => m.id === "static-no-context")?.contextWindow).toBeUndefined();
   });
 
   test("provider context-window caps apply to stale cached metadata", async () => {
diff --git a/tests/context-cap-unknown-window.test.ts b/tests/context-cap-unknown-window.test.ts
new file mode 100644
index 0000000000..79e46ff410
--- /dev/null
+++ b/tests/context-cap-unknown-window.test.ts
@@ -0,0 +1,25 @@
+import { describe, expect, test } from "bun:test";
+import {
+  applyProviderContextCap,
+  resolveUnknownRoutedContextWindow,
+} from "../src/providers/context-cap";
+
+describe("unknown routed context windows", () => {
+  test("an enabled cap fills a missing window instead of inventing 128k", () => {
+    expect(resolveUnknownRoutedContextWindow(350_000)).toBe(350_000);
+    expect(applyProviderContextCap(undefined, 350_000)).toBeUndefined();
+  });
+
+  test("a discovered window is only lowered", () => {
+    expect(applyProviderContextCap(128_000, 350_000)).toBe(128_000);
+    expect(applyProviderContextCap(500_000, 350_000)).toBe(350_000);
+  });
+
+  test("no cap keeps the conservative 128k fallback", () => {
+    expect(resolveUnknownRoutedContextWindow(undefined)).toBe(128_000);
+  });
+
+  test("a fractional cap that floors to zero does not invent a zero window", () => {
+    expect(resolveUnknownRoutedContextWindow(0.5)).toBe(128_000);
+  });
+});