fix: recognize common multimodal text models

backend/internal/app/model_capability.go

@@ -207,9 +207,13 @@ func legacyImageSizeValues() []string {
 }
 
 func DefaultModelCapabilityConfigForModel(protocol string, modelName string) *ModelCapabilityConfig {
-	// 文本模型是否支持视觉输入不能从协议或模型名可靠推断，默认关闭，由管理员按真实上游能力开启。
 	streaming := true
-	text := &TextCapabilityConfig{Streaming: &streaming, References: TextReferenceConfig{PromptMaxChars: 32000}}
+	textReferences := TextReferenceConfig{PromptMaxChars: 32000}
+	if isKnownMultimodalTextModel(protocol, modelName) {
+		textReferences.MaxImages = 16
+		textReferences.MaxImageBytes = 30 * 1024 * 1024
+	}
+	text := &TextCapabilityConfig{Streaming: &streaming, References: textReferences}
 	video := &VideoCapabilityConfig{
 		References:        VideoReferenceConfig{PromptMaxChars: DefaultVideoPromptMaxChars, MinImages: 0, MaxImages: 9, MaxImageBytes: 30 * 1024 * 1024, MaxVideos: 0, MaxVideoBytes: 0, MaxVideoDuration: 0, MaxAudios: 0, MaxAudioBytes: 0, MaxAudioDuration: 0},
 		Duration:          VideoDurationConfig{Selection: "range", Min: 1, Max: 15, Step: 1, Default: 6},
@@ -275,6 +279,11 @@ func DefaultModelCapabilityConfigForModel(protocol string, modelName string) *Mo
 	return &ModelCapabilityConfig{Version: 1, Text: text, Image: DefaultImageCapabilityConfig(protocol, modelName), Video: video}
 }
 
+func isKnownMultimodalTextModel(protocol string, modelName string) bool {
+	model := strings.ToLower(strings.TrimSpace(modelName))
+	return strings.Contains(model, "gemini") || strings.Contains(model, "gpt") || strings.Contains(model, "doubao") || strings.Contains(model, "豆包") || strings.EqualFold(strings.TrimSpace(protocol), "gemini")
+}
+
 func DecodeModelCapabilityConfig(raw string) (*ModelCapabilityConfig, error) {
 	if strings.TrimSpace(raw) == "" {
 		return nil, nil
@@ -323,6 +332,14 @@ func NormalizeModelCapabilityConfigForModel(capability string, protocol string,
 			return nil, BadAuthRequest("请配置文本模型能力参数")
 		}
 		text := *input.Text
+		if isKnownMultimodalTextModel(protocol, modelName) && text.References.MaxImages == 0 {
+			// 旧配置是在文本模型默认不接收图片时生成的；Gemini 文本模型
+			// 通过兼容协议仍然会接收图片 parts，因此自动修复旧的 0 上限。
+			text.References.MaxImages = 16
+			if text.References.MaxImageBytes == 0 {
+				text.References.MaxImageBytes = 30 * 1024 * 1024
+			}
+		}
 		if text.Streaming == nil {
 			streaming := true
 			text.Streaming = &streaming

backend/internal/app/model_capability_test.go

@@ -6,6 +6,23 @@ import (
 	"testing"
 )
 
+func TestDefaultModelCapabilityConfigRecognizesKnownTextVisionFamilies(t *testing.T) {
+	gemini := DefaultModelCapabilityConfigForModel("chat-completion", "gemini-3.8-flash-high")
+	gpt := DefaultModelCapabilityConfigForModel("chat-completion", "gpt-4o")
+	doubao := DefaultModelCapabilityConfigForModel("chat-completion", "doubao-seed-1.6-flash")
+	plain := DefaultModelCapabilityConfigForModel("chat-completion", "plain-text-model")
+
+	if gemini.Text == nil || gemini.Text.References.MaxImages != 16 || gemini.Text.References.MaxImageBytes != 30*1024*1024 {
+		t.Fatalf("Gemini text vision defaults = %#v", gemini.Text)
+	}
+	if gpt.Text == nil || gpt.Text.References.MaxImages != 16 || doubao.Text == nil || doubao.Text.References.MaxImages != 16 {
+		t.Fatalf("known multimodal text vision defaults: gpt=%#v doubao=%#v", gpt.Text, doubao.Text)
+	}
+	if plain.Text == nil || plain.Text.References.MaxImages != 0 {
+		t.Fatalf("plain text vision defaults = %#v", plain.Text)
+	}
+}
+
 func TestValidateImageTaskRejectsOversizedGrokPromptByUTF8Bytes(t *testing.T) {
 	prompt := strings.Repeat("中", 4001)
 	input := canvasGenerationInput{

web/src/lib/model-capabilities.ts

@@ -105,6 +105,14 @@ export const STANDARD_IMAGE_SIZE_VALUES = [
     "1024x1536",
 ] as const;
 
+const DEFAULT_TEXT_REFERENCE_MAX_IMAGES = 16;
+
+function isKnownMultimodalTextModel(protocol?: ModelProtocol, model = "") {
+    // These model families currently expose vision through their text/chat
+    // endpoint, including when they are wrapped by an OpenAI-compatible gateway.
+    return /gemini|gpt|doubao|豆包/i.test(model) || protocol === "gemini";
+}
+
 export function normalizeCapabilityString(value: string) {
     const normalized = value.trim();
     return normalized.startsWith("string:") ? normalized.slice("string:".length) : normalized;
@@ -286,8 +294,13 @@ export function defaultImageCapabilityConfig(protocol?: ModelProtocol, model = "
 export function defaultModelCapabilityConfig(protocol?: ModelProtocol, model = ""): ModelCapabilityConfig {
     const text: TextCapabilityConfig = {
         streaming: true,
-        // 文本模型的视觉能力必须由管理员明确开启，不能根据模型名猜测。
-        references: { promptMaxChars: 32000, maxImages: 0, maxImageBytes: 0, maxVideos: 0, maxVideoBytes: 0 },
+        references: {
+            promptMaxChars: 32000,
+            maxImages: isKnownMultimodalTextModel(protocol, model) ? DEFAULT_TEXT_REFERENCE_MAX_IMAGES : 0,
+            maxImageBytes: isKnownMultimodalTextModel(protocol, model) ? 30 * 1024 * 1024 : 0,
+            maxVideos: 0,
+            maxVideoBytes: 0,
+        },
     };
     const video: VideoCapabilityConfig = {
         references: {
@@ -398,7 +411,10 @@ export function modelCapabilityConfigFor(config: { channels: Array<{ id: string;
     const fallback = defaultModelCapabilityConfig(cost?.protocol, modelName);
     if (!cost?.capabilityConfig) return fallback;
     const capabilityConfig = normalizeModelCapabilityConfig(cost.capabilityConfig);
-    const text = capabilityConfig.text ? { ...fallback.text!, ...capabilityConfig.text, references: { ...fallback.text!.references, ...capabilityConfig.text.references } } : fallback.text;
+    const mergedText = capabilityConfig.text ? { ...fallback.text!, ...capabilityConfig.text, references: { ...fallback.text!.references, ...capabilityConfig.text.references } } : fallback.text;
+    const text = mergedText && isKnownMultimodalTextModel(cost?.protocol, modelName) && mergedText.references.maxImages === 0
+        ? { ...mergedText, references: { ...mergedText.references, maxImages: DEFAULT_TEXT_REFERENCE_MAX_IMAGES, maxImageBytes: mergedText.references.maxImageBytes || 30 * 1024 * 1024 } }
+        : mergedText;
     const video = capabilityConfig.video ? { ...fallback.video!, ...capabilityConfig.video, references: { ...fallback.video!.references, ...capabilityConfig.video.references } } : fallback.video;
     const configuredImage = capabilityConfig.image;
     const image = configuredImage

web/test/model-capabilities.test.ts

@@ -31,3 +31,16 @@ test("raising the video default leaves text and image limits untouched", () => {
     assert.equal(profile.text!.references.promptMaxChars, 32000);
     assert.equal(profile.image!.references.promptMaxChars, 32000);
 });
+
+test("known multimodal text model families default to image input", () => {
+    const gemini = defaultModelCapabilityConfig("chat-completion", "gemini-3.8-flash-high").text!;
+    const gpt = defaultModelCapabilityConfig("chat-completion", "gpt-4o").text!;
+    const doubao = defaultModelCapabilityConfig("chat-completion", "doubao-seed-1.6-flash").text!;
+    const plain = defaultModelCapabilityConfig("chat-completion", "plain-text-model").text!;
+
+    assert.equal(gemini.references.maxImages, 16);
+    assert.equal(gemini.references.maxImageBytes, 30 * 1024 * 1024);
+    assert.equal(gpt.references.maxImages, 16);
+    assert.equal(doubao.references.maxImages, 16);
+    assert.equal(plain.references.maxImages, 0);
+});