feat(canvas): AI 拆图 - 完善规划、图层编辑与素材复用

backend/internal/app/provider_layer_planning_reference_test.go

@@ -0,0 +1,35 @@
+package app
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func TestLayerPlanningDeliversOwnedPixelsToVisionModel(t *testing.T) {
+	s, _, pixels := cloudAgentVisionFixture(t)
+	input := canvasGenerationInput{
+		Mode: "text", Prompt: "规划图层", Metadata: map[string]interface{}{"edit": "layer-planning"},
+		ReferenceImages: []providerMedia{{StorageKey: "resource:ref-one", URL: "https://invalid.example/stale.png", DataURL: "data:image/png;base64,c3RhbGU="}},
+	}
+	if err := s.hydrateGenerationMedia("user", &input, providerMediaHydrationPolicy{preferURL: true}); err != nil {
+		t.Fatal(err)
+	}
+	want := "data:image/png;base64," + base64.StdEncoding.EncodeToString(pixels)
+	if input.ReferenceImages[0].DataURL != want || input.ReferenceImages[0].URL != "" {
+		t.Fatal("planning did not use the authorized resource's actual pixels")
+	}
+	content, err := textChatContent(input)
+	if err != nil {
+		t.Fatal(err)
+	}
+	raw, err := json.Marshal(content)
+	if err != nil || !strings.Contains(string(raw), want) || strings.Contains(string(raw), "invalid.example") {
+		t.Fatalf("vision payload lost stored pixels: %v", err)
+	}
+	foreign := canvasGenerationInput{Mode: "text", Metadata: input.Metadata, ReferenceImages: []providerMedia{{StorageKey: "resource:ref-one"}}}
+	if err := s.hydrateGenerationMedia("other-user", &foreign, providerMediaHydrationPolicy{preferURL: true}); err == nil {
+		t.Fatal("planning allowed another user's resource")
+	}
+}

backend/internal/app/provider_media_hydration.go

@@ -19,8 +19,10 @@ func (s *Service) hydrateGenerationMedia(userID string, input *canvasGenerationI
 	groups := [][]providerMedia{input.ReferenceImages, input.ReferenceVideos, input.ReferenceAudios}
 	for groupIndex, group := range groups {
 		mediaPolicy := policy
-		// 仅 Agent 图片走内存字节；视频、音频及普通生成任务保留原来的协议策略。
-		if groupIndex == 0 && input.Mode == "text" && input.AgentRequests != nil && input.AgentRequests.Canonical != nil {
+		// 识图规划与 Agent 看图都发送归属校验后的实际像素，避免供应商无法读取签名 URL。
+		// 视频、音频及其他普通生成任务保留原来的协议策略。
+		inlineVision := input.Mode == "text" && ((input.AgentRequests != nil && input.AgentRequests.Canonical != nil) || metadataString(input.Metadata, "edit") == "layer-planning")
+		if groupIndex == 0 && inlineVision {
 			mediaPolicy = providerMediaHydrationPolicy{imageOnly: true}
 			if input.Config.CapabilityConfig != nil && input.Config.CapabilityConfig.Text != nil {
 				limits := input.Config.CapabilityConfig.Text.References
@@ -107,6 +109,9 @@ func (s *Service) hydrateProviderMedia(userID string, media *providerMedia, poli
 	}
 	mimeType := normalizedMediaMimeType(firstNonEmpty(media.MimeType, resource.MimeType), data)
 	media.DataURL = dataURL(mimeType, data)
+	if policy.imageOnly {
+		media.URL = ""
+	}
 	media.MimeType = mimeType
 	media.Bytes = int64(len(data))
 	media.Width = resource.Width

backend/internal/protocol/image_tools_test.go

@@ -36,6 +36,9 @@ func TestImageToolsProvidersBuildAsyncImageRequests(t *testing.T) {
 			if body["image"] != "https://cdn.example/source.png" {
 				t.Fatalf("source image mapping = %#v", body["image"])
 			}
+			if tc.provider == "image-tools-layer-decomposition" && body["output_format"] != "png" {
+				t.Fatalf("layer outputs must preserve alpha: %#v", body["output_format"])
+			}
 			result, err := adapter.ParseCreate(context.Background(), []byte(`{"data":{"id":"task-image-tool-1","status":"processing"}}`))
 			if err != nil || result.TaskID != "task-image-tool-1" || result.Status != StatusProcessing {
 				t.Fatalf("create result = %#v, err = %v", result, err)