feat(plugins): 方舟 Agent Plan - 合并 PR #527 图片视频插件并阻止图片误用 asset://

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/internal/app/channel_models.go

@@ -438,7 +438,7 @@ func (s *Service) syncLogicalModelsFromChannelModel(actor *model.User, channelMo
 }
 
 func supportsTokenBilling(capability string, protocol model.ChannelInterfaceType) bool {
-	return capability == "text" || (capability == "video" && protocol == model.ChannelInterfaceVolcengineArkVideo)
+	return capability == "text" || (capability == "video" && model.IsVolcengineArkVideoProtocol(protocol))
 }
 
 func (s *Service) normalizeChannelModelPriceTiers(req ChannelModelRequest, capability string, protocol model.ChannelInterfaceType, fallbackProviderModelKey string) ([]model.ChannelModelPriceTier, error) {

backend/internal/app/model_capability.go

@@ -153,7 +153,7 @@ func DefaultImageCapabilityConfig(protocol string, modelName string) *ImageCapab
 		image.ResponseFormat = ParameterSupport{Supported: true}
 		image.OutputFormat = ParameterSupport{Supported: false}
 		image.MaxOutputs = 1
-	case model.ChannelInterfaceVolcengineArkImage:
+	case model.ChannelInterfaceVolcengineArkImage, model.ChannelInterfaceVolcengineArkAgentPlanImage:
 		image.References.MaskSupported = false
 		image.Quality.Supported = false
 		image.TransparentBackground.Supported = false
@@ -229,7 +229,7 @@ func DefaultModelCapabilityConfigForModel(protocol string, modelName string) *Mo
 	case model.ChannelInterfaceGeminiVeo:
 		video.Duration = VideoDurationConfig{Selection: "enum", Values: []int{4, 6, 8}, Default: 6}
 		video.Resolutions = []string{"720p", "1080p"}
-	case model.ChannelInterfaceVolcengineArkVideo:
+	case model.ChannelInterfaceVolcengineArkVideo, model.ChannelInterfaceVolcengineArkAgentPlanVideo:
 		video.Operations = append(video.Operations, "reference_to_video", "audio_to_video")
 		video.References.MaxVideos, video.References.MaxAudios = 3, 3
 		video.References.MaxVideoBytes, video.References.MaxAudioBytes = 200*1024*1024, 15*1024*1024
@@ -798,7 +798,7 @@ func validateVideoTask(profile *VideoCapabilityConfig, input canvasGenerationInp
 	if len(input.ReferenceImages) > profile.References.MaxImages || len(input.ReferenceVideos) > profile.References.MaxVideos || len(input.ReferenceAudios) > profile.References.MaxAudios {
 		return BadAuthRequest("参考素材数量超过当前模型限制")
 	}
-	if input.Config.InterfaceType == string(model.ChannelInterfaceVolcengineArkVideo) && len(input.ReferenceAudios) > 0 && len(input.ReferenceImages) == 0 && len(input.ReferenceVideos) == 0 {
+	if model.IsVolcengineArkVideoProtocol(model.ChannelInterfaceType(input.Config.InterfaceType)) && len(input.ReferenceAudios) > 0 && len(input.ReferenceImages) == 0 && len(input.ReferenceVideos) == 0 {
 		return BadAuthRequest("火山方舟全模态参考不支持纯音频或文本+音频，请同时添加参考图片或参考视频")
 	}
 	if len(input.ReferenceImages) < profile.References.MinImages {

backend/internal/app/price_validation.go

@@ -14,7 +14,7 @@ func ValidateChannelModelPrice(billingMode string, capability string, protocol m
 		return unitPrice >= 0
 	case "token":
 		if capability == "video" {
-			return protocol == model.ChannelInterfaceVolcengineArkVideo && inputPrice >= 0 && outputPrice >= 0 && cachedPrice >= 0
+			return model.IsVolcengineArkVideoProtocol(protocol) && inputPrice >= 0 && outputPrice >= 0 && cachedPrice >= 0
 		}
 		if capability != "" && capability != "text" {
 			return false

backend/internal/app/provider.go

@@ -464,7 +464,7 @@ type providerMediaHydrationPolicy struct {
 func providerMediaHydrationPolicyFor(ctx context.Context, input canvasGenerationInput) providerMediaHydrationPolicy {
 	policy := providerMediaHydrationPolicy{preferURL: providerPrefersMediaURLs(input.Config.InterfaceType, input)}
 	switch strings.TrimSpace(input.Config.InterfaceType) {
-	case string(model.ChannelInterfaceNewAPIVideo), string(model.ChannelInterfaceNewAPIChannel1), string(model.ChannelInterfaceNewAPIChannel2), string(model.ChannelInterfaceVolcengineArkVideo), string(model.ChannelInterfaceMiniMaxVideo):
+	case string(model.ChannelInterfaceNewAPIVideo), string(model.ChannelInterfaceNewAPIChannel1), string(model.ChannelInterfaceNewAPIChannel2), string(model.ChannelInterfaceVolcengineArkVideo), string(model.ChannelInterfaceVolcengineArkAgentPlanVideo), string(model.ChannelInterfaceMiniMaxVideo):
 		policy.requireURL = true
 		policy.preferURL = true
 	}
@@ -488,11 +488,11 @@ func providerPrefersMediaURLs(interfaceType string, input canvasGenerationInput)
 	}
 	switch strings.TrimSpace(interfaceType) {
 	case string(model.ChannelInterfaceChatCompletion), string(model.ChannelInterfaceOpenAIResponse), string(model.ChannelInterfaceClaudeAPI),
-		string(model.ChannelInterfaceGrokImage), string(model.ChannelInterfaceVolcengineArkImage),
+		string(model.ChannelInterfaceGrokImage), string(model.ChannelInterfaceVolcengineArkImage), string(model.ChannelInterfaceVolcengineArkAgentPlanImage),
 		string(model.ChannelInterfaceXAIVideo), string(model.ChannelInterfaceNovitaVideo),
 		string(model.ChannelInterfaceMiniMaxVideo), string(model.ChannelInterfaceNewAPIVideo),
 		string(model.ChannelInterfaceNewAPIChannel1), string(model.ChannelInterfaceNewAPIChannel2),
-		string(model.ChannelInterfaceVolcengineArkVideo):
+		string(model.ChannelInterfaceVolcengineArkVideo), string(model.ChannelInterfaceVolcengineArkAgentPlanVideo):
 		return true
 	}
 	if isGrokVideoConfig(input.Config) || isSeedanceVideoConfig(input.Config) || isArkPlanVideoConfig(input.Config) {

backend/internal/app/provider_ark_private_assets.go

@@ -47,7 +47,8 @@ func withoutProviderAnalytics(ctx context.Context) context.Context {
 }
 
 func (s *Service) prepareArkPrivateAssetReferences(ctx context.Context, userID string, input *canvasGenerationInput) error {
-	if input == nil || !isArkPrivateAssetVideoConfig(input.Config) || !parseBool(input.Config.ArkPrivateAssetUpload, true) {
+	// 可信素材 asset:// 仅方舟视频协议支持；Agent Plan Seedream 等图片渠道不能上传或改写。
+	if input == nil || input.Mode != "video" || !isArkPrivateAssetVideoConfig(input.Config) || !parseBool(input.Config.ArkPrivateAssetUpload, true) {
 		return nil
 	}
 	hasOwnedReference := false
@@ -107,7 +108,12 @@ func (s *Service) prepareArkPrivateAssetReferences(ctx context.Context, userID s
 }
 
 func isArkPrivateAssetVideoConfig(config providerConfig) bool {
-	return config.InterfaceType == string(model.ChannelInterfaceVolcengineArkVideo) || isArkPlanVideoConfig(config)
+	iface := strings.TrimSpace(config.InterfaceType)
+	// Agent Plan 图片与视频共用 /api/plan/v3；图片协议不得进入可信素材同步。
+	if iface == string(model.ChannelInterfaceVolcengineArkImage) || iface == "volcengine-ark-agent-plan-image" {
+		return false
+	}
+	return iface == string(model.ChannelInterfaceVolcengineArkVideo) || iface == "volcengine-ark-agent-plan-video" || isArkPlanVideoConfig(config)
 }
 
 func arkPrivateAssetAutomaticSyncEnabled(setting arkPrivateAssetSettingValue) (bool, error) {

backend/internal/app/provider_audio.go

@@ -172,7 +172,7 @@ func asyncAudioErrorMessage(state map[string]interface{}) string {
 func decodeProviderDataURL(value string) (string, []byte, error) {
 	header, encoded, ok := strings.Cut(value, ",")
 	if !ok || !strings.HasPrefix(header, "data:") || !strings.HasSuffix(strings.ToLower(header), ";base64") {
-		return "", nil, errors.New("音频 data URL 格式无效")
+		return "", nil, errors.New("data URL 格式无效")
 	}
 	mimeType := strings.TrimSuffix(strings.TrimPrefix(header, "data:"), ";base64")
 	data, err := base64.StdEncoding.DecodeString(encoded)

backend/internal/app/provider_image.go

@@ -28,7 +28,7 @@ func runImageTask(ctx context.Context, input canvasGenerationInput) (map[string]
 	if input.Config.InterfaceType == string(model.ChannelInterfaceVolcengineJiMengImage) {
 		return runVolcengineJiMengImageTask(ctx, input)
 	}
-	if input.Config.InterfaceType == string(model.ChannelInterfaceVolcengineArkImage) {
+	if model.IsVolcengineArkImageProtocol(model.ChannelInterfaceType(input.Config.InterfaceType)) {
 		return runVolcengineArkImageTask(ctx, input)
 	}
 	if input.Config.InterfaceType == string(model.ChannelInterfaceGeminiImage) {

backend/internal/app/provider_task_cancellation.go

@@ -194,15 +194,15 @@ func (s *Service) providerCancellationInput(task *model.Task) (canvasGenerationI
 }
 
 func supportsProviderCancellation(interfaceType string) bool {
-	return interfaceType == string(model.ChannelInterfaceGeminiVeo) || interfaceType == string(model.ChannelInterfaceVolcengineArkVideo)
+	return interfaceType == string(model.ChannelInterfaceGeminiVeo) || model.IsVolcengineArkVideoProtocol(model.ChannelInterfaceType(interfaceType))
 }
 
 func cancelProviderTask(ctx context.Context, config providerConfig, providerRequestID string) error {
-	switch config.InterfaceType {
-	case string(model.ChannelInterfaceGeminiVeo):
+	switch {
+	case config.InterfaceType == string(model.ChannelInterfaceGeminiVeo):
 		path := "/" + strings.TrimLeft(providerRequestID, "/") + ":cancel"
 		return postGeminiJSON(ctx, config, path, map[string]any{}, &map[string]any{})
-	case string(model.ChannelInterfaceVolcengineArkVideo):
+	case model.IsVolcengineArkVideoProtocol(model.ChannelInterfaceType(config.InterfaceType)):
 		path := "/contents/generations/tasks/" + url.PathEscape(providerRequestID)
 		return deleteProviderTask(ctx, config, path)
 	default:
@@ -228,8 +228,8 @@ func deleteProviderTask(ctx context.Context, config providerConfig, path string)
 }
 
 func queryProviderCancellation(ctx context.Context, config providerConfig, providerRequestID string) (providerCancellationOutcome, string, error) {
-	switch config.InterfaceType {
-	case string(model.ChannelInterfaceGeminiVeo):
+	switch {
+	case config.InterfaceType == string(model.ChannelInterfaceGeminiVeo):
 		// Gemini 用 operation.error 表示取消终态，此处必须保留该字段再判断，
 		// 不能让通用 JSON 解包提前把它转换成请求失败。
 		var operation geminiOperation
@@ -248,7 +248,7 @@ func queryProviderCancellation(ctx context.Context, config providerConfig, provi
 			return providerCancellationFailed, firstNonEmpty(message, "failed"), nil
 		}
 		return providerCancellationSucceeded, "succeeded", nil
-	case string(model.ChannelInterfaceVolcengineArkVideo):
+	case model.IsVolcengineArkVideoProtocol(model.ChannelInterfaceType(config.InterfaceType)):
 		var state map[string]any
 		if err := getJSON(ctx, config, "/contents/generations/tasks/"+url.PathEscape(providerRequestID), &state); err != nil {
 			return "", "", err

backend/internal/app/provider_test.go

@@ -2224,6 +2224,48 @@ func TestArkPlanConfigStaysSeparateFromSeedanceVideosEndpoint(t *testing.T) {
 	}
 }
 
+func TestArkPlanImageConfigDoesNotUseVideoAssetPath(t *testing.T) {
+	imageConfig := providerConfig{
+		InterfaceType: "volcengine-ark-agent-plan-image",
+		BaseURL:       "https://ark.cn-beijing.volces.com/api/plan/v3",
+		Model:         "doubao-seedream-4-0-250828",
+	}
+	if isArkPlanVideoConfig(imageConfig) {
+		t.Fatal("agent plan image must not match isArkPlanVideoConfig")
+	}
+	if isArkPrivateAssetVideoConfig(imageConfig) {
+		t.Fatal("agent plan image must not trigger ark private asset sync")
+	}
+	if isSeedanceVideoConfig(imageConfig) {
+		t.Fatal("agent plan image must not match isSeedanceVideoConfig")
+	}
+
+	videoConfig := providerConfig{
+		InterfaceType: "volcengine-ark-agent-plan-video",
+		BaseURL:       "https://ark.cn-beijing.volces.com/api/plan/v3",
+		Model:         "doubao-seedance-1-5-pro-251215",
+	}
+	if !isArkPlanVideoConfig(videoConfig) || !isArkPrivateAssetVideoConfig(videoConfig) {
+		t.Fatal("agent plan video should keep video/asset path")
+	}
+}
+
+func TestPrepareArkPrivateAssetReferencesSkipsImageMode(t *testing.T) {
+	svc := &Service{}
+	err := svc.prepareArkPrivateAssetReferences(context.Background(), "user-1", &canvasGenerationInput{
+		Mode: "image",
+		Config: providerConfig{
+			InterfaceType:         "volcengine-ark-agent-plan-image",
+			BaseURL:               "https://ark.cn-beijing.volces.com/api/plan/v3",
+			ArkPrivateAssetUpload: "true",
+		},
+		ReferenceImages: []providerMedia{{StorageKey: "resource:res-1", URL: "https://example.com/ref.png"}},
+	})
+	if err != nil {
+		t.Fatalf("prepareArkPrivateAssetReferences() error = %v", err)
+	}
+}
+
 func TestVolcengineArkVideoProtocolUsesContentTaskAndDownloadsResult(t *testing.T) {
 	t.Setenv("CANVAS_ALLOW_PRIVATE_UPSTREAMS", "true")
 	paths := make([]string, 0, 3)

backend/internal/app/provider_video.go

@@ -287,7 +287,7 @@ func runSeedanceVideosTask(ctx context.Context, input canvasGenerationInput, pol
 
 func runSeedanceAgentPlanVideoTask(ctx context.Context, input canvasGenerationInput, pollPolicy videoPollPolicy) (map[string]interface{}, error) {
 	providerName := "Seedance"
-	if input.Config.InterfaceType == string(model.ChannelInterfaceVolcengineArkVideo) {
+	if model.IsVolcengineArkVideoProtocol(model.ChannelInterfaceType(input.Config.InterfaceType)) {
 		providerName = "火山方舟"
 	}
 	// Agent Plan 与 Videos API 使用相同的恢复合同：已有任务只查询，不重复创建。
@@ -299,7 +299,7 @@ func runSeedanceAgentPlanVideoTask(ctx context.Context, input canvasGenerationIn
 		if err != nil {
 			return nil, err
 		}
-		if input.Config.InterfaceType == string(model.ChannelInterfaceVolcengineArkVideo) {
+		if model.IsVolcengineArkVideoProtocol(model.ChannelInterfaceType(input.Config.InterfaceType)) {
 			for _, item := range content {
 				if item["type"] == "image_url" {
 					item["role"] = "reference_image"

backend/internal/app/provider_video_options.go

@@ -5,6 +5,8 @@ import (
 	"strconv"
 	"strings"
 	"time"
+
+	"infinite-canvas/backend/internal/model"
 )
 
 func isPublicMediaURL(value string) bool {
@@ -13,8 +15,8 @@ func isPublicMediaURL(value string) bool {
 }
 
 func isSeedanceVideoConfig(config providerConfig) bool {
-	model := strings.ToLower(config.Model)
-	return strings.Contains(model, "seedance") || strings.Contains(model, "doubao-seedance") || isArkPlanVideoConfig(config)
+	modelName := strings.ToLower(config.Model)
+	return strings.Contains(modelName, "seedance") || strings.Contains(modelName, "doubao-seedance") || isArkPlanVideoConfig(config)
 }
 
 func isGrokVideoConfig(config providerConfig) bool {
@@ -22,7 +24,15 @@ func isGrokVideoConfig(config providerConfig) bool {
 }
 
 func isArkPlanVideoConfig(config providerConfig) bool {
-	return strings.Contains(strings.ToLower(config.BaseURL), "/api/plan/v3")
+	if !strings.Contains(strings.ToLower(config.BaseURL), "/api/plan/v3") {
+		return false
+	}
+	// Agent Plan 图片与视频共用 /api/plan/v3；按协议排除图片，避免 Seedream 误走视频/可信素材路径。
+	iface := strings.TrimSpace(config.InterfaceType)
+	if iface == string(model.ChannelInterfaceVolcengineArkImage) || iface == "volcengine-ark-agent-plan-image" {
+		return false
+	}
+	return true
 }
 
 func normalizeImageQuality(value string) string {

backend/internal/generation/registry.go

@@ -70,7 +70,7 @@ func OfficialDeclarativeVideoInterface(interfaceType string) (string, bool) {
 		return "NewAPI 媒体任务", true
 	case string(model.ChannelInterfaceXAIVideo):
 		return "xAI", true
-	case string(model.ChannelInterfaceVolcengineArkVideo):
+	case string(model.ChannelInterfaceVolcengineArkVideo), string(model.ChannelInterfaceVolcengineArkAgentPlanVideo):
 		return "火山方舟", true
 	case string(model.ChannelInterfaceVolcengineJiMengVideo):
 		return "即梦", true
@@ -91,6 +91,8 @@ func OfficialDeclarativeImageInterface(interfaceType string) (string, bool) {
 		return "Grok Images", true
 	case string(model.ChannelInterfaceVolcengineArkImage):
 		return "火山方舟图片", true
+	case string(model.ChannelInterfaceVolcengineArkAgentPlanImage):
+		return "火山方舟 Agent Plan 图片", true
 	case string(model.ChannelInterfaceVolcengineJiMengImage):
 		return "即梦图片", true
 	default:

backend/internal/handler/security.go

@@ -244,11 +244,11 @@ func interfaceAllowsProxyPath(interfaceType model.ChannelInterfaceType, requestP
 		return requestPath == "/messages"
 	case model.ChannelInterfaceOpenAIImage, model.ChannelInterfaceGrokImage:
 		return requestPath == "/images/generations" || requestPath == "/images/edits"
-	case model.ChannelInterfaceVolcengineArkImage:
+	case model.ChannelInterfaceVolcengineArkImage, model.ChannelInterfaceVolcengineArkAgentPlanImage:
 		return requestPath == "/images/generations"
 	case model.ChannelInterfaceOpenAIAudio:
 		return requestPath == "/audio/speech"
-	case model.ChannelInterfaceAsyncAudio, model.ChannelInterfaceNewAPIVideo, model.ChannelInterfaceNewAPIChannel1, model.ChannelInterfaceNewAPIChannel2, model.ChannelInterfaceXAIVideo, model.ChannelInterfaceVolcengineArkVideo, model.ChannelInterfaceVolcengineJiMengImage, model.ChannelInterfaceVolcengineJiMengVideo, model.ChannelInterfaceGeminiVeo, model.ChannelInterfaceGeminiImage, model.ChannelInterfaceNovitaVideo, model.ChannelInterfaceMiniMaxVideo:
+	case model.ChannelInterfaceAsyncAudio, model.ChannelInterfaceNewAPIVideo, model.ChannelInterfaceNewAPIChannel1, model.ChannelInterfaceNewAPIChannel2, model.ChannelInterfaceXAIVideo, model.ChannelInterfaceVolcengineArkVideo, model.ChannelInterfaceVolcengineArkAgentPlanVideo, model.ChannelInterfaceVolcengineJiMengImage, model.ChannelInterfaceVolcengineJiMengVideo, model.ChannelInterfaceGeminiVeo, model.ChannelInterfaceGeminiImage, model.ChannelInterfaceNovitaVideo, model.ChannelInterfaceMiniMaxVideo:
 		return false
 	default:
 		return true

backend/internal/model/models.go

@@ -63,16 +63,18 @@ const (
 	ChannelInterfaceClaudeAPI             ChannelInterfaceType = "claude-api"
 	ChannelInterfaceOpenAIImage           ChannelInterfaceType = "openai-image"
 	ChannelInterfaceGrokImage             ChannelInterfaceType = "grok-image"
-	ChannelInterfaceVolcengineArkImage    ChannelInterfaceType = "volcengine-ark-image"
-	ChannelInterfaceVolcengineJiMengImage ChannelInterfaceType = "volcengine-jimeng-image"
-	ChannelInterfaceGeminiImage           ChannelInterfaceType = "gemini-image"
-	ChannelInterfaceOpenAIAudio           ChannelInterfaceType = "openai-audio"
-	ChannelInterfaceAsyncAudio            ChannelInterfaceType = "async-audio"
-	ChannelInterfaceNewAPIVideo           ChannelInterfaceType = "newapi"
-	ChannelInterfaceNewAPIChannel1        ChannelInterfaceType = "newapi-channel-1"
-	ChannelInterfaceNewAPIChannel2        ChannelInterfaceType = "newapi-channel-2"
-	ChannelInterfaceXAIVideo              ChannelInterfaceType = "xai-video"
-	ChannelInterfaceVolcengineArkVideo    ChannelInterfaceType = "volcengine-ark-video"
+	ChannelInterfaceVolcengineArkImage           ChannelInterfaceType = "volcengine-ark-image"
+	ChannelInterfaceVolcengineArkAgentPlanImage  ChannelInterfaceType = "volcengine-ark-agent-plan-image"
+	ChannelInterfaceVolcengineJiMengImage        ChannelInterfaceType = "volcengine-jimeng-image"
+	ChannelInterfaceGeminiImage                  ChannelInterfaceType = "gemini-image"
+	ChannelInterfaceOpenAIAudio                  ChannelInterfaceType = "openai-audio"
+	ChannelInterfaceAsyncAudio                   ChannelInterfaceType = "async-audio"
+	ChannelInterfaceNewAPIVideo                  ChannelInterfaceType = "newapi"
+	ChannelInterfaceNewAPIChannel1               ChannelInterfaceType = "newapi-channel-1"
+	ChannelInterfaceNewAPIChannel2               ChannelInterfaceType = "newapi-channel-2"
+	ChannelInterfaceXAIVideo                     ChannelInterfaceType = "xai-video"
+	ChannelInterfaceVolcengineArkVideo           ChannelInterfaceType = "volcengine-ark-video"
+	ChannelInterfaceVolcengineArkAgentPlanVideo  ChannelInterfaceType = "volcengine-ark-agent-plan-video"
 	ChannelInterfaceVolcengineJiMengVideo ChannelInterfaceType = "volcengine-jimeng-video"
 	ChannelInterfaceGeminiVeo             ChannelInterfaceType = "gemini-veo"
 	ChannelInterfaceNovitaVideo           ChannelInterfaceType = "novita-video"

backend/internal/model/volcengine_ark.go

@@ -0,0 +1,11 @@
+package model
+
+// IsVolcengineArkImageProtocol 识别官方 Ark 图片与 Agent Plan 图片协议。
+func IsVolcengineArkImageProtocol(protocol ChannelInterfaceType) bool {
+	return protocol == ChannelInterfaceVolcengineArkImage || protocol == ChannelInterfaceVolcengineArkAgentPlanImage
+}
+
+// IsVolcengineArkVideoProtocol 识别官方 Ark 视频与 Agent Plan 视频协议。
+func IsVolcengineArkVideoProtocol(protocol ChannelInterfaceType) bool {
+	return protocol == ChannelInterfaceVolcengineArkVideo || protocol == ChannelInterfaceVolcengineArkAgentPlanVideo
+}

backend/internal/model/volcengine_ark_test.go

@@ -0,0 +1,24 @@
+package model
+
+import "testing"
+
+func TestIsVolcengineArkProtocols(t *testing.T) {
+	if !IsVolcengineArkImageProtocol(ChannelInterfaceVolcengineArkImage) {
+		t.Fatal("official ark image should match")
+	}
+	if !IsVolcengineArkImageProtocol(ChannelInterfaceVolcengineArkAgentPlanImage) {
+		t.Fatal("agent plan image should match")
+	}
+	if IsVolcengineArkImageProtocol(ChannelInterfaceOpenAIImage) {
+		t.Fatal("openai image should not match ark image")
+	}
+	if !IsVolcengineArkVideoProtocol(ChannelInterfaceVolcengineArkVideo) {
+		t.Fatal("official ark video should match")
+	}
+	if !IsVolcengineArkVideoProtocol(ChannelInterfaceVolcengineArkAgentPlanVideo) {
+		t.Fatal("agent plan video should match")
+	}
+	if IsVolcengineArkVideoProtocol(ChannelInterfaceNewAPIVideo) {
+		t.Fatal("newapi video should not match ark video")
+	}
+}

backend/internal/protocol/docs/volcengine-ark-agent-plan-image.md

@@ -0,0 +1,37 @@
+# 火山方舟 Agent Plan 图片生成
+
+本插件面向火山方舟 Agent Plan 专属图片入口。请求体与官方 Ark Seedream 协议一致，但 Base URL 与 API Key 必须来自 Agent Plan 控制台。
+
+## 接口、鉴权与区域
+
+{{OPERATIONS}}
+
+```http
+POST {channel_base_url}/api/plan/v3/images/generations
+Authorization: Bearer <AGENT_PLAN_API_KEY>
+Content-Type: application/json
+```
+
+与官方 `/api/v3` 的差异：
+
+| 项 | 官方 Ark | Agent Plan |
+| --- | --- | --- |
+| Base URL | `https://ark.cn-beijing.volces.com/api/v3` | `https://ark.cn-beijing.volces.com/api/plan/v3` |
+| API Key | 推理接入控制台 Key | Agent Plan 专属 Key |
+| 计费 | 推理后付费/项目额度 | AFP 套餐抵扣 |
+| 请求体 | Seedream images/generations | 同左 |
+
+不要把官方 Key 打到 `/api/plan/v3`，也不要把 Agent Plan Key 打到 `/api/v3`。
+
+## 参数与字段映射
+
+{{PARAMETERS}}
+
+当前实现：`aspectRatio -> size`，所有 `images -> image[]`；`extra` 可透传 `size`、`sequential_image_generation`、`sequential_image_generation_options`、`watermark`。
+
+## 官方资料
+
+- [火山方舟 Agent Plan 文档](https://www.volcengine.com/docs/82379/2375486)
+- [火山方舟视觉模型 API 文档](https://www.volcengine.com/docs/82379)
+
+{{CONTRACT}}

backend/internal/protocol/docs/volcengine-ark-agent-plan-video.md

@@ -0,0 +1,43 @@
+# 火山方舟 Agent Plan 视频任务
+
+本插件使用 Agent Plan 专属 Contents Generations Tasks 异步接口。提示词、参考图、参考视频和参考音频组成 `content[]`；比例、分辨率、时长、生成音频和水印位于顶层任务参数。请求体与官方 Ark Seedance 一致，入口改为 `/api/plan/v3`。
+
+## 接口与鉴权
+
+{{OPERATIONS}}
+
+```http
+POST {channel_base_url}/api/plan/v3/contents/generations/tasks
+GET  {channel_base_url}/api/plan/v3/contents/generations/tasks/{task_id}
+DELETE {channel_base_url}/api/plan/v3/contents/generations/tasks/{task_id}
+Authorization: Bearer <AGENT_PLAN_API_KEY>
+Content-Type: application/json
+```
+
+与官方 `/api/v3` 的差异：
+
+| 项 | 官方 Ark | Agent Plan |
+| --- | --- | --- |
+| Base URL | `https://ark.cn-beijing.volces.com/api/v3` | `https://ark.cn-beijing.volces.com/api/plan/v3` |
+| API Key | 推理接入控制台 Key | Agent Plan 专属 Key |
+| 计费 | 推理后付费/项目额度 | AFP 套餐抵扣 |
+| 请求体 | contents/generations/tasks | 同左 |
+
+## 参数与字段映射
+
+{{PARAMETERS}}
+
+全模态参考的素材映射如下：
+
+- 图片：`{"type":"image_url","image_url":{"url":"..."},"role":"reference_image"}`
+- 视频：`{"type":"video_url","video_url":{"url":"..."},"role":"reference_video"}`
+- 音频：`{"type":"audio_url","audio_url":{"url":"..."},"role":"reference_audio"}`
+
+单次最多发送 9 张图片、3 个视频和 3 个音频。参考素材可以组合使用，但上游不支持纯音频或“文本 + 音频”。
+
+## 官方资料
+
+- [火山方舟 Agent Plan 文档](https://www.volcengine.com/docs/82379/2375486)
+- [火山方舟内容生成任务文档](https://www.volcengine.com/docs/82379/1520757)
+
+{{CONTRACT}}

backend/internal/protocol/manifest.go

@@ -711,6 +711,8 @@ func mediaReferencesFromManifestValue(value any, kind string, ephemeral bool) []
 			if reference.DataURL == "" {
 				reference.DataURL = firstString(typed, "data_url", "b64_json")
 			}
+			// OpenAI / Ark 等渠道常直接返回裸 b64_json；声明式结果下载要求 data URL。
+			reference.DataURL = normalizeManifestInlineDataURL(reference.DataURL, reference.Kind, firstString(typed, "output_format", "mime_type", "mimeType"))
 			if reference.URL != "" || reference.DataURL != "" {
 				result = append(result, reference)
 			}
@@ -719,6 +721,50 @@ func mediaReferencesFromManifestValue(value any, kind string, ephemeral bool) []
 	return result
 }
 
+func normalizeManifestInlineDataURL(value, kind, formatHint string) string {
+	value = strings.TrimSpace(value)
+	if value == "" || strings.HasPrefix(value, "data:") {
+		return value
+	}
+	return "data:" + manifestInlineMediaMIME(kind, formatHint) + ";base64," + value
+}
+
+func manifestInlineMediaMIME(kind, formatHint string) string {
+	hint := strings.ToLower(strings.TrimSpace(formatHint))
+	switch hint {
+	case "image/png", "image/jpeg", "image/webp", "image/gif", "audio/mpeg", "audio/wav", "audio/ogg", "video/mp4", "video/webm":
+		return hint
+	case "image/jpg":
+		return "image/jpeg"
+	case "audio/mp3":
+		return "audio/mpeg"
+	case "png":
+		return "image/png"
+	case "jpeg", "jpg":
+		return "image/jpeg"
+	case "webp":
+		return "image/webp"
+	case "gif":
+		return "image/gif"
+	case "mp3", "mpeg":
+		return "audio/mpeg"
+	case "wav":
+		return "audio/wav"
+	case "mp4":
+		return "video/mp4"
+	case "webm":
+		return "video/webm"
+	}
+	switch strings.ToLower(strings.TrimSpace(kind)) {
+	case "audio":
+		return "audio/mpeg"
+	case "video":
+		return "video/mp4"
+	default:
+		return "image/png"
+	}
+}
+
 var (
 	manifestMDImageRegex   = regexp.MustCompile(`!\[[^\]]*\]\(([^)\s]+)\)`)
 	manifestHTMLVideoRegex = regexp.MustCompile(`(?i)<video[^>]+src=['"]([^'"]+)['"]`)

backend/internal/protocol/official_catalog_test.go

@@ -599,6 +599,36 @@ func TestOfficialArkSeedreamMapsAspectRatioToPixelSize(t *testing.T) {
 	}
 }
 
+func TestOfficialArkAgentPlanPluginsUsePlanPaths(t *testing.T) {
+	image := officialPackageAdapter(t, "volcengine-ark-agent-plan-seedream.yingce-plugin", "volcengine-ark-agent-plan-image")
+	imageCreate, err := image.BuildCreate(context.Background(), RequestContext{Request: GenerationRequest{Model: "doubao-seedream-5-0-260128", Prompt: "circle", AspectRatio: "1:1"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if imageCreate.Path != "/api/plan/v3/images/generations" {
+		t.Fatalf("agent plan image create = %#v", imageCreate)
+	}
+	if body := manifestTestBody(t, imageCreate); body["size"] != "2048x2048" {
+		t.Fatalf("agent plan image size = %#v", body["size"])
+	}
+
+	video := officialPackageAdapter(t, "volcengine-ark-agent-plan-seedance.yingce-plugin", "volcengine-ark-agent-plan-video")
+	videoCreate, err := video.BuildCreate(context.Background(), RequestContext{Request: GenerationRequest{Model: "doubao-seedance-2-0-260128", Prompt: "walk", AspectRatio: "16:9", Resolution: "720p", Duration: 5}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if videoCreate.Path != "/api/plan/v3/contents/generations/tasks" {
+		t.Fatalf("agent plan video create = %#v", videoCreate)
+	}
+	poll, err := video.BuildPoll(context.Background(), PollContext{TaskID: "task-1", Model: "doubao-seedance-2-0-260128"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if poll.Path != "/api/plan/v3/contents/generations/tasks/task-1" {
+		t.Fatalf("agent plan video poll = %#v", poll)
+	}
+}
+
 func TestOfficialGeminiImageMapsQualityToImageSize(t *testing.T) {
 	adapter := officialPackageAdapter(t, "google-gemini-image.yingce-plugin", "gemini-image")
 	tests := []struct {
@@ -754,3 +784,27 @@ func TestOfficialOpenAIAudioSpeedDefaultsInvalidAndZeroValues(t *testing.T) {
 		})
 	}
 }
+
+func TestOfficialArkSeedreamParsesB64JSONAsDataURL(t *testing.T) {
+	for _, tc := range []struct {
+		packageName, providerID string
+	}{
+		{"volcengine-ark-seedream.yingce-plugin", "volcengine-ark-image"},
+		{"volcengine-ark-agent-plan-seedream.yingce-plugin", "volcengine-ark-agent-plan-image"},
+	} {
+		t.Run(tc.providerID, func(t *testing.T) {
+			adapter := officialPackageAdapter(t, tc.packageName, tc.providerID)
+			result, err := adapter.ParseCreate(context.Background(), []byte(`{"created":1,"data":[{"b64_json":"aW1hZ2U=","output_format":"jpeg","size":"1824x1024"}]}`))
+			if err != nil {
+				t.Fatal(err)
+			}
+			if result.Result == nil || len(result.Result.Images) != 1 {
+				t.Fatalf("result = %#v", result.Result)
+			}
+			dataURL := result.Result.Images[0].DataURL
+			if dataURL != "data:image/jpeg;base64,aW1hZ2U=" {
+				t.Fatalf("DataURL = %q, want jpeg data URL from b64_json + output_format", dataURL)
+			}
+		})
+	}
+}