feat(combos): enroll text-only members in vision sidecar from the UI

docs-site/src/content/docs/guides/combos.md

@@ -319,6 +319,16 @@ even when every target supports images — the catalog drops `image` from `input
 image-bearing requests are rejected with HTTP 400 before any target is called. `"auto"` (or
 omitting the field) keeps the automatic intersection.
 
+In the dashboard, the **Image / multimodal** switch can be enabled whenever every target is a
+known catalog row. Members that advertise image input stay unchanged; members that do not are
+automatically declared text-only on save (`modelCapabilities[model].inputModalities = ["text"]`)
+so the [Vision Sidecar](/guides/sidecars/) describes their images. The switch's hint names the
+members that will be enrolled. Turning the switch off disables image input for the combo but
+keeps those provider declarations, and `PUT /api/combos` accepts the same enrollment as an
+optional top-level `visionSidecarTargets` array of exact `{ provider, model }` targets
+(request-only; it is never stored on the combo and is rejected while `imageInput` is
+`"disabled"` when the enrollment list is non-empty).
+
 ## Encrypted v2 sub-agent tasks
 
 There is one important limitation for Codex v2 sub-agents ([issue #92](https://github.com/lidge-jun/opencodex/issues/92)).

docs-site/src/content/docs/guides/sidecars.md

@@ -133,6 +133,9 @@ allow attachments instead of blocking them before the sidecar runs. When
 use the `gpt-5.6-luna` fallback. Startup still migrates an explicitly persisted legacy
 `gpt-5.4-mini` value to `gpt-5.6-luna`; that migration applies to a stored value, not to an absent
 model field.
+Saving a multimodal combo from the dashboard enrolls its non-image members automatically: the
+Combos page sends them as `visionSidecarTargets` on `PUT /api/combos`, which writes the exact
+text-only declaration on the member's provider, so no manual config edit and reload is needed.
 The first-party DeepSeek `deepseek-flash` model is native multimodal (`text` and `image`) and does
 not use this sidecar by default. Explicit `noVisionModels` or text-only declarations remain
 authoritative. First-party `deepseek-chat`, `deepseek-reasoner`, and `deepseek-v4-flash` remain

gui/src/combo-capabilities.ts

@@ -1,14 +1,47 @@
 import type { ComboTarget } from "./combo-workspace-data";
 import type { ModelOption } from "./components/combo-workspace-types";
 
-/** Whether every selected target advertises image input (incomplete rows fail closed). */
+type ComboImageMemberKind = "vision" | "sidecar" | "missing";
+
+/**
+ * One combo member's image story. A row advertising image is either natively
+ * multimodal or an already-declared Vision Sidecar consumer (the catalog widens
+ * both). A catalog row without image can be declared text-only on save so the
+ * sidecar covers it. Rows absent from the catalog fail closed.
+ */
+function imageMemberKind(target: ComboTarget, models: ModelOption[]): ComboImageMemberKind {
+  const provider = target.provider.trim();
+  const modelId = target.model.trim();
+  if (!provider || !modelId) return "missing";
+  const model = models.find((row) => row.provider === provider && row.id === modelId);
+  if (!model) return "missing";
+  return model.inputModalities?.includes("image") ? "vision" : "sidecar";
+}
+
+/** Whether images can be enabled: every target is known and either images natively or can be declared text-only. */
 export function comboImagesSupported(targets: ComboTarget[], models: ModelOption[]): boolean {
   if (targets.length === 0) return false;
-  return targets.every((target) => {
+  return targets.every((target) => imageMemberKind(target, models) !== "missing");
+}
+
+/**
+ * Exact targets that need a text-only declaration so the Vision Sidecar covers
+ * them when the combo accepts images. Deduplicated in submission order.
+ */
+export function comboVisionSidecarTargets(
+  targets: ComboTarget[],
+  models: ModelOption[],
+): Array<{ provider: string; model: string }> {
+  const seen = new Set<string>();
+  const out: Array<{ provider: string; model: string }> = [];
+  for (const target of targets) {
+    if (imageMemberKind(target, models) !== "sidecar") continue;
     const provider = target.provider.trim();
-    const modelId = target.model.trim();
-    if (!provider || !modelId) return false;
-    const model = models.find((row) => row.provider === provider && row.id === modelId);
-    return !!model?.inputModalities?.includes("image");
-  });
+    const model = target.model.trim();
+    const key = `${provider}/${model}`;
+    if (seen.has(key)) continue;
+    seen.add(key);
+    out.push({ provider, model });
+  }
+  return out;
 }

gui/src/combo-workspace-data.ts

@@ -412,9 +412,13 @@ export function draftEquals(a: ComboItem, b: ComboItem): boolean {
   });
 }
 
-export function toPutBody(item: ComboItem, options: { renameFrom?: string } = {}): {
+export function toPutBody(
+  item: ComboItem,
+  options: { renameFrom?: string; visionSidecarTargets?: Array<{ provider: string; model: string }> } = {},
+): {
   id: string;
   renameFrom?: string;
+  visionSidecarTargets?: Array<{ provider: string; model: string }>;
   combo: {
     targets: ComboTarget[];
     strategy: ComboStrategy;
@@ -431,6 +435,7 @@ export function toPutBody(item: ComboItem, options: { renameFrom?: string } = {}
   return {
     id: item.id.trim(),
     ...(options.renameFrom ? { renameFrom: options.renameFrom } : {}),
+    ...(options.visionSidecarTargets?.length ? { visionSidecarTargets: options.visionSidecarTargets } : {}),
     combo: {
       targets: item.targets.map((target) => weighted
         ? { provider: target.provider.trim(), model: target.model.trim(), weight: target.weight ?? 1 }

gui/src/components/combo-workspace-controls.tsx

@@ -1,6 +1,6 @@
 import { useState } from "react";
 import type { ComboEffort, ComboStrategy, ComboTarget, ProviderQuotaStates } from "../combo-workspace-data";
-import { comboImagesSupported } from "../combo-capabilities";
+import { comboImagesSupported, comboVisionSidecarTargets } from "../combo-capabilities";
 import { COMBO_EFFORTS, COMBO_STRATEGIES, COMBO_STRATEGY_LABEL_KEYS, newComboTarget } from "../combo-workspace-data";
 import { IconArrowDown, IconArrowUp, IconGrip, IconPlus, IconTrash } from "../icons";
 import { useT } from "../i18n/shared";
@@ -102,6 +102,14 @@ export function ComboCapabilities({
   const imagesSupported = comboImagesSupported(targets, models);
   // Default: checked (auto) when supported; force off when any target lacks image.
   const effectiveOn = imagesSupported && imageInput !== "disabled";
+  const sidecarTargets = comboVisionSidecarTargets(targets, models);
+  const imageHint = !imagesSupported
+    ? t("cws.capability.imageInputUnavailable")
+    : sidecarTargets.length > 0
+      ? t("cws.capability.imageInputSidecarHint", {
+        models: sidecarTargets.map(({ provider, model }) => `${provider}/${model}`).join(", "),
+      })
+      : t("cws.capability.imageInputHint");
 
   return (
     <section className="cwi-capabilities" aria-label={t("cws.capabilities")}>
@@ -110,7 +118,7 @@ export function ComboCapabilities({
         <div>
           <span className="cwi-capability-label">{t("cws.capability.imageInput")}</span>
           <p className="muted cwi-capability-hint">
-            {imagesSupported ? t("cws.capability.imageInputHint") : t("cws.capability.imageInputUnavailable")}
+            {imageHint}
           </p>
         </div>
         <Switch

gui/src/i18n/de.ts

@@ -2682,8 +2682,9 @@ export const de: Record<TKey, string> = {
   "cws.field.defaultEffort": "Standard-Reasoning",
   "cws.field.defaultEffortNone": "Keine (Ziel-Standard)",
   "cws.field.defaultEffortHint": "Nur verwendet, wenn der Client keinen Reasoning-Aufwand sendet. Optionen sind die Schnittmenge der beworbenen Aufwände der gewählten Ziele.",
-  "cws.capability.imageInputUnavailable": "Erst verfügbar, wenn jedes gewählte Ziel Bildeingabe unterstützt.",
+  "cws.capability.imageInputUnavailable": "Wähle zuerst alle Ziele aus dem Katalog — unbekannte Modelle kann der Vision Sidecar nicht abdecken.",
   "cws.capability.imageInputHint": "Standardmäßig aktiv, wenn jedes Ziel Bilder unterstützt. Ausschalten für nur Text.",
+  "cws.capability.imageInputSidecarHint": "Standardmäßig aktiv. {models} wird beim Speichern als text-only deklariert und nutzt den Vision Sidecar für Bilder.",
   "cws.capability.imageInput": "Bild / multimodal",
   "cws.capability.adaptiveEffort": "Adaptive Denkstufen",
   "cws.capability.adaptiveEffortHint": "Aus: Ziele ohne Denkstufen-Regelung blenden die Auswahl für die gesamte Kombination aus. An: Solche Ziele bleiben nutzbar, und die Auswahl zeigt weiterhin die Stufen der übrigen Ziele.",

gui/src/i18n/en.ts

@@ -2782,8 +2782,9 @@ export const en = {
   "cws.field.defaultEffort": "Default reasoning",
   "cws.field.defaultEffortNone": "None (target default)",
   "cws.field.defaultEffortHint": "Used only when the client omits reasoning effort. Options are the intersection of the selected targets' advertised efforts; targets without catalog effort metadata offer none.",
-  "cws.capability.imageInputUnavailable": "Unavailable until every selected target supports image input.",
+  "cws.capability.imageInputUnavailable": "Pick every target from the catalog first — unknown models can't be covered by the Vision Sidecar.",
   "cws.capability.imageInputHint": "On by default when every target supports images. Turn off to accept text only.",
+  "cws.capability.imageInputSidecarHint": "On by default. {models} will be declared text-only on save and use the Vision Sidecar for images.",
   "cws.capability.imageInput": "Image / multimodal",
   "cws.capability.adaptiveEffort": "Adaptive reasoning ladder",
   "cws.capability.adaptiveEffortHint": "Off: a target with no reasoning control hides the effort picker for the whole combo. On: those targets stay usable and the picker keeps the levels the remaining targets share.",

gui/src/i18n/fr.ts

@@ -2667,8 +2667,9 @@ export const fr: Record<TKey, string> = {
   "cws.emptyTitle": "Créer votre première combinaison",
   "cws.empty.createDesc": "Nommez un modèle virtuel et enchaînez au moins deux services principaux.",
   "cws.backToAll": "Retour à toutes les combinaisons",
-  "cws.capability.imageInputUnavailable": "Indisponible tant que toutes les cibles sélectionnées ne prennent pas en charge les images.",
-  "cws.capability.imageInputHint": "Activé par défaut lorsque toutes les cibles prennent en charge les images. Désactivez cette option pour n’accepter que du texte.",
+  "cws.capability.imageInputUnavailable": "Sélectionnez d'abord chaque cible dans le catalogue — les modèles inconnus ne peuvent pas être couverts par le Vision Sidecar.",
+  "cws.capability.imageInputHint": "Activé par défaut lorsque toutes les cibles prennent en charge les images. Désactivez cette option pour n'accepter que du texte.",
+  "cws.capability.imageInputSidecarHint": "Activé par défaut. Lors de l’enregistrement, les cibles {models} seront déclarées en mode texte uniquement et utiliseront le Vision Sidecar pour les images.",
   "cws.capability.imageInput": "Images / multimodal",
   "cws.capability.adaptiveEffort": "Échelle de raisonnement adaptative",
   "cws.capability.adaptiveEffortHint": "Désactivé : une cible sans réglage de raisonnement masque le sélecteur pour toute la combinaison. Activé : ces cibles restent utilisables et le sélecteur conserve les niveaux communs aux autres cibles.",

gui/src/i18n/ja.ts

@@ -2734,8 +2734,9 @@ export const ja: Record<TKey, string> = {
   "cws.field.defaultEffort": "デフォルトの推論",
   "cws.field.defaultEffortNone": "なし(ターゲットのデフォルト)",
   "cws.field.defaultEffortHint": "クライアントが推論負荷を省略した場合のみ使用されます。選択肢は選択ターゲットが広告する負荷の交差です。",
-  "cws.capability.imageInputUnavailable": "選択した全ターゲットが画像入力に対応すると有効になります。",
+  "cws.capability.imageInputUnavailable": "まずすべてのターゲットをカタログから選択してください。不明なモデルはVision Sidecarでカバーできません。",
   "cws.capability.imageInputHint": "全ターゲットが画像対応なら既定でオン。オフにするとテキストのみ。",
+  "cws.capability.imageInputSidecarHint": "既定でオン。{models} は保存時にテキスト専用として宣言され、画像にはVision Sidecarを使用します。",
   "cws.capability.imageInput": "画像 / マルチモーダル",
   "cws.capability.adaptiveEffort": "適応的な推論レベル",
   "cws.capability.adaptiveEffortHint": "オフ: 推論レベルを持たない対象があると、コンボ全体のセレクターが消えます。オン: その対象はそのまま使え、セレクターには残りの対象で共通するレベルが表示されます。",

gui/src/i18n/ko.ts

@@ -2721,8 +2721,9 @@ export const ko: Record<TKey, string> = {
   "cws.field.defaultEffort": "기본 추론 수준",
   "cws.field.defaultEffortNone": "없음 (대상 기본값)",
   "cws.field.defaultEffortHint": "클라이언트가 추론 수준을 생략한 경우에만 사용합니다. 옵션은 선택한 대상이 광고하는 수준의 교집합입니다.",
-  "cws.capability.imageInputUnavailable": "선택한 모든 대상이 이미지 입력을 지원해야 사용할 수 있습니다.",
+  "cws.capability.imageInputUnavailable": "먼저 모든 대상을 카탈로그에서 선택하세요. 알 수 없는 모델은 Vision Sidecar로 처리할 수 없습니다.",
   "cws.capability.imageInputHint": "모든 대상이 이미지를 지원하면 기본으로 켜집니다. 끄면 텍스트만 허용합니다.",
+  "cws.capability.imageInputSidecarHint": "기본으로 켜짐. {models}은(는) 저장 시 텍스트 전용으로 선언되어 이미지에 Vision Sidecar를 사용합니다.",
   "cws.capability.imageInput": "이미지 / 멀티모달",
   "cws.capability.adaptiveEffort": "적응형 추론 단계",
   "cws.capability.adaptiveEffortHint": "끔: 추론 단계를 조절할 수 없는 대상이 하나라도 있으면 콤보 전체의 선택기가 사라집니다. 켬: 그런 대상도 그대로 쓰면서, 선택기에는 나머지 대상이 공통으로 지원하는 단계가 남습니다.",

gui/src/i18n/ru.ts

@@ -2805,8 +2805,9 @@ export const ru: Record<TKey, string> = {
   "cws.field.defaultEffort": "Рассуждения по умолчанию",
   "cws.field.defaultEffortNone": "Нет (по умолчанию для цели)",
   "cws.field.defaultEffortHint": "Используется, только если клиент не указал уровень рассуждений. Варианты — пересечение заявленных уровней выбранных целей.",
-  "cws.capability.imageInputUnavailable": "Доступно, когда все выбранные цели поддерживают ввод изображений.",
+  "cws.capability.imageInputUnavailable": "Сначала выберите все цели из каталога — неизвестные модели нельзя обработать через Vision Sidecar.",
   "cws.capability.imageInputHint": "Включено по умолчанию, если все цели поддерживают изображения. Выключите, чтобы принимать только текст.",
+  "cws.capability.imageInputSidecarHint": "Включено по умолчанию. При сохранении для {models} будет указана только текстовая модальность; изображения будут обрабатываться через Vision Sidecar.",
   "cws.capability.imageInput": "Изображения / мультимодальность",
   "cws.capability.adaptiveEffort": "Адаптивная шкала рассуждений",
   "cws.capability.adaptiveEffortHint": "Выкл.: цель без настройки рассуждений скрывает выбор уровня для всей комбинации. Вкл.: такие цели остаются доступными, а в выборе сохраняются уровни, общие для остальных целей.",

gui/src/i18n/tr.ts

@@ -2724,8 +2724,9 @@ export const tr: Record<TKey, string> = {
   "cws.field.defaultEffort": "Varsayılan akıl yürütme",
   "cws.field.defaultEffortNone": "Yok (hedef varsayılanı)",
   "cws.field.defaultEffortHint": "Yalnızca istemci akıl yürütme çabasını belirtmediğinde (atladığında) kullanılır. Seçenekler, seçilen hedeflerin duyurulan çabalarının kesişimidir; katalog çaba meta verisi olmayan hedefler hiçbir seçenek sunmaz.",
-  "cws.capability.imageInputUnavailable": "Seçilen tüm hedefler görsel girişini destekleyene kadar kullanılamaz.",
+  "cws.capability.imageInputUnavailable": "Önce tüm hedefleri katalogdan seçin — bilinmeyen modeller Vision Sidecar ile kapsanamaz.",
   "cws.capability.imageInputHint": "Tüm hedefler görselleri desteklediğinde varsayılan olarak açıktır. Yalnızca metin kabul etmek için kapatın.",
+  "cws.capability.imageInputSidecarHint": "Varsayılan olarak açık. {models} kaydedildiğinde yalnızca metin olarak bildirilir ve görseller için Vision Sidecar kullanır.",
   "cws.capability.imageInput": "Görsel / çok modlu",
   "cws.capability.adaptiveEffort": "Uyarlanabilir akıl yürütme düzeyi",
   "cws.capability.adaptiveEffortHint": "Kapalı: akıl yürütme denetimi olmayan bir hedef, tüm kombinasyonun seçicisini gizler. Açık: bu hedefler kullanılabilir kalır ve seçici, kalan hedeflerin ortak düzeylerini gösterir.",

gui/src/i18n/zh-TW.ts

@@ -2077,8 +2077,9 @@ export const zhTW: Record<TKey, string> = {
   "cws.field.defaultEffort": "預設推理級別",
   "cws.field.defaultEffortNone": "無（使用目標預設）",
   "cws.field.defaultEffortHint": "僅在客戶端未指定推理級別時使用。客戶端值優先，每個目標會按自身能力進行處理。",
-  "cws.capability.imageInputUnavailable": "所有已選目標都支援圖片輸入後才可使用。",
+  "cws.capability.imageInputUnavailable": "請先從目錄選擇所有目標——未知模型無法由 Vision Sidecar 處理。",
   "cws.capability.imageInputHint": "所有目標都支援圖片時預設開啟；關閉後僅接受文字。",
+  "cws.capability.imageInputSidecarHint": "預設開啟。{models} 將在儲存時宣告為純文字，並使用 Vision Sidecar 處理圖片。",
   "cws.capability.imageInput": "圖片 / 多模態",
   "cws.capability.adaptiveEffort": "自適應推理層級",
   "cws.capability.adaptiveEffortHint": "關閉：只要有一個目標不支援推理層級，整個組合的選擇器都會消失。開啟：這些目標仍可使用，選擇器保留其餘目標共有的層級。",

gui/src/i18n/zh.ts

@@ -2702,8 +2702,9 @@ export const zh: Record<TKey, string> = {
   "cws.field.defaultEffort": "默认推理级别",
   "cws.field.defaultEffortNone": "无（使用目标默认）",
   "cws.field.defaultEffortHint": "仅在客户端未指定推理级别时使用。选项为所选目标已公布努力级别的交集。",
-  "cws.capability.imageInputUnavailable": "所有已选目标均支持图片输入后才可用。",
+  "cws.capability.imageInputUnavailable": "请先从目录中选择所有目标——未知模型无法由 Vision Sidecar 处理。",
   "cws.capability.imageInputHint": "所有目标均支持图片时默认开启；关闭后仅接受文本。",
+  "cws.capability.imageInputSidecarHint": "默认开启。{models} 将在保存时声明为纯文本，并使用 Vision Sidecar 处理图片。",
   "cws.capability.imageInput": "图片 / 多模态",
   "cws.capability.adaptiveEffort": "自适应推理档位",
   "cws.capability.adaptiveEffortHint": "关闭：只要有一个目标不支持推理档位，整个组合的选择器都会消失。开启：这些目标仍可使用，选择器保留其余目标共有的档位。",