feat(piagent): bypass planner for direct image generation modes

For image / pure-image / image-generator / image-action modes, skip the
task-planning round trip and hand the prompt straight to the image model,
deriving the requested image count from the composer settings. Also
normalize completed image-generator nodes back into regular image nodes
(clearing the image-generator layer role and defaulting Tone to "visual")
when merging generated results and when restoring finished nodes, so a
finished generator node is not re-appended as a duplicate.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-07-09 01:09:59 +08:00
parent a39e18950c
commit ffd8d78df7
5 changed files with 573 additions and 60 deletions
+158 -36
View File
@@ -547,11 +547,12 @@ func (s *DesignService) AgentChat(ctx context.Context, projectID string, req des
actionID := newID()
targetNodeID := imageGeneratorTargetNodeID(req.Messages)
isolatedImageGenerator := isImageGeneratorRequest(mode, targetNodeID)
directImageMode := isDirectImageGenerationMode(mode, isolatedImageGenerator)
enableWebSearch := req.EnableWebSearch || toolConfigEnablesSearch(req.ToolConfig)
memory := buildAgentMemory(projectScopedToAgentThread(project, threadID), prompt, mode, req.Messages)
if !isolatedImageGenerator {
project.Messages = append(project.Messages,
design.Message{
nextMessages := []design.Message{
{
ID: newID(),
Role: "user",
Type: "user",
@@ -562,8 +563,11 @@ func (s *DesignService) AgentChat(ctx context.Context, projectID string, req des
StepID: newID(),
CreatedAt: now,
},
agentPlanningMessage(threadID, actionID, prompt, enableWebSearch, now.Add(time.Millisecond)),
)
}
if !directImageMode {
nextMessages = append(nextMessages, agentPlanningMessage(threadID, actionID, prompt, enableWebSearch, now.Add(time.Millisecond)))
}
project.Messages = append(project.Messages, nextMessages...)
project.Status = design.StatusExploring
project.LastThreadID = threadID
project.UpdatedAt = now
@@ -571,31 +575,38 @@ func (s *DesignService) AgentChat(ctx context.Context, projectID string, req des
if err := s.repo.Save(ctx, project); err != nil {
return design.AgentThread{}, err
}
s.streamAgentPlanningMessage(ctx, project.ID, threadID, design.AgentTaskPlanRequest{
Project: projectScopedToAgentThread(projectWithoutGeneratingNodes(project), threadID),
Prompt: prompt,
Mode: mode,
Messages: req.Messages,
Memory: buildAgentMemory(projectScopedToAgentThread(project, threadID), prompt, mode, req.Messages),
WebSearchEnabled: enableWebSearch,
})
project, err = s.repo.Get(ctx, project.ID)
if err != nil {
return design.AgentThread{}, err
}
memory = buildAgentMemory(projectScopedToAgentThread(project, threadID), prompt, mode, req.Messages)
if cancelled, err := s.projectThreadCancelled(ctx, project.ID, threadID); err != nil {
return design.AgentThread{}, err
} else if cancelled {
return design.AgentThread{
ThreadID: threadID,
Status: "cancelled",
ThreadIDType: threadIDType,
Project: project,
}, nil
if !directImageMode {
s.streamAgentPlanningMessage(ctx, project.ID, threadID, design.AgentTaskPlanRequest{
Project: projectScopedToAgentThread(projectWithoutGeneratingNodes(project), threadID),
Prompt: prompt,
Mode: mode,
Messages: req.Messages,
Memory: buildAgentMemory(projectScopedToAgentThread(project, threadID), prompt, mode, req.Messages),
WebSearchEnabled: enableWebSearch,
})
project, err = s.repo.Get(ctx, project.ID)
if err != nil {
return design.AgentThread{}, err
}
memory = buildAgentMemory(projectScopedToAgentThread(project, threadID), prompt, mode, req.Messages)
if cancelled, err := s.projectThreadCancelled(ctx, project.ID, threadID); err != nil {
return design.AgentThread{}, err
} else if cancelled {
return design.AgentThread{
ThreadID: threadID,
Status: "cancelled",
ThreadIDType: threadIDType,
Project: project,
}, nil
}
}
}
taskPlan := s.planAgentTask(ctx, projectScopedToAgentThread(projectWithoutGeneratingNodes(project), threadID), prompt, mode, req.Messages, enableWebSearch, isolatedImageGenerator, memory)
var taskPlan design.AgentTaskPlan
if directImageMode {
taskPlan = directImageTaskPlan(prompt, req.Messages)
} else {
taskPlan = s.planAgentTask(ctx, projectScopedToAgentThread(projectWithoutGeneratingNodes(project), threadID), prompt, mode, req.Messages, enableWebSearch, isolatedImageGenerator, memory)
}
if cancelled, err := s.projectThreadCancelled(ctx, project.ID, threadID); err != nil {
return design.AgentThread{}, err
} else if cancelled {
@@ -606,7 +617,7 @@ func (s *DesignService) AgentChat(ctx context.Context, projectID string, req des
Project: project,
}, nil
}
if !isolatedImageGenerator {
if !isolatedImageGenerator && !directImageMode {
if err := s.maybeAddPlannedResearchStep(ctx, project.ID, threadID, projectScopedToAgentThread(project, threadID), prompt, mode, req.Messages, enableWebSearch, taskPlan); err != nil {
return design.AgentThread{}, err
}
@@ -647,7 +658,9 @@ func (s *DesignService) AgentChat(ctx context.Context, projectID string, req des
assistantMessageID := ""
if !isolatedImageGenerator {
if shouldGenerateImage {
applyPlannerTaskPlanMessage(&project, threadID, prompt, taskPlan)
if !directImageMode {
applyPlannerTaskPlanMessage(&project, threadID, prompt, taskPlan)
}
} else {
assistantMessageID = preparePlannerForChatMessage(&project, threadID, actionID, prompt, now.Add(time.Millisecond))
}
@@ -3468,6 +3481,84 @@ func shouldGenerateForAgentPrompt(prompt string, mode string, messages []design.
return false
}
func isDirectImageGenerationMode(mode string, isolated bool) bool {
if isolated {
return true
}
modeText := strings.ToLower(strings.TrimSpace(mode))
return modeText == "image" || modeText == "pure-image" || modeText == "image-generator" || strings.Contains(modeText, "image-action")
}
func directImageTaskPlan(prompt string, messages []design.AgentChatMessage) design.AgentTaskPlan {
count := imageModeSettingsCount(messages)
if count <= 0 {
count = 1
}
if count > 10 {
count = 10
}
tasks := make([]design.AgentImageTask, 0, count)
for index := 0; index < count; index++ {
tasks = append(tasks, design.AgentImageTask{
Title: generatedTaskFallbackTitle(prompt, index, count),
Brief: prompt,
})
}
return design.AgentTaskPlan{
Intent: "image",
Reason: "direct image mode",
ImageCount: count,
ImageTasks: tasks,
}
}
func imageModeSettingsCount(messages []design.AgentChatMessage) int {
for _, message := range messages {
for _, content := range message.Contents {
if !strings.EqualFold(strings.TrimSpace(content.TextSource), "settings") {
continue
}
if count := imageModeSettingsCountFromText(content.Text); count > 0 {
return count
}
}
}
return 0
}
func imageModeSettingsCountFromText(text string) int {
lower := strings.ToLower(text)
for _, marker := range []string{"数量", "count"} {
index := strings.Index(lower, marker)
if index < 0 {
continue
}
if count := firstPositiveInt(lower[index+len(marker):]); count > 0 {
return count
}
}
return 0
}
func firstPositiveInt(text string) int {
value := 0
seen := false
for _, char := range text {
if char < '0' || char > '9' {
if seen {
break
}
continue
}
seen = true
value = value*10 + int(char-'0')
if value > 10 {
return 10
}
}
return value
}
func shouldForceImageMode(mode string) bool {
modeText := strings.ToLower(strings.TrimSpace(mode))
return strings.Contains(modeText, "image-action") || modeText == "image" || modeText == "pure-image"
@@ -4048,9 +4139,6 @@ func (s *DesignService) clearStaleGeneratingPlaceholders(project design.Project,
func (s *DesignService) restoreFinishedGeneratedImageNodes(project design.Project, now time.Time) (design.Project, bool) {
artifacts := generatedImageArtifacts(project.Messages)
if len(artifacts) == 0 {
return project, false
}
artifactsByID := generatedImageArtifactsByID(project.Messages)
usedArtifacts := make(map[string]bool)
next := make([]design.Node, 0, len(project.Nodes))
@@ -4068,6 +4156,10 @@ func (s *DesignService) restoreFinishedGeneratedImageNodes(project design.Projec
}
changed = true
}
if normalized, ok := completedImageGeneratorNode(node); ok {
node = normalized
changed = true
}
if project.Status == design.StatusReady && node.Type == design.NodeTypeImage && node.Status == "generating" && !isGeneratedImageContent(node.Content) {
changed = true
continue
@@ -4155,6 +4247,17 @@ func restoreNodeFromGeneratedArtifact(node design.Node, artifact design.Generato
return node
}
func completedImageGeneratorNode(node design.Node) (design.Node, bool) {
if node.Type != design.NodeTypeImage || node.LayerRole != "image-generator" || node.Status != "success" || !isGeneratedImageContent(node.Content) {
return node, false
}
node.LayerRole = ""
if strings.TrimSpace(node.Tone) == "" || node.Tone == "image-generator" {
node.Tone = "visual"
}
return node, true
}
func generationPlaceholdersAreStale(project design.Project, now time.Time) bool {
if !hasGeneratingCanvasWork(project.Nodes) {
return false
@@ -4361,20 +4464,24 @@ func imageGeneratorTargetNodeID(messages []design.AgentChatMessage) string {
func mergeGeneratedNodesIntoTarget(nodes []design.Node, generated []design.Node, targetNodeID string) []design.Node {
if strings.TrimSpace(targetNodeID) == "" || len(generated) == 0 {
return append(nodes, generated...)
return append(nodes, completedImageGeneratorNodes(generated)...)
}
targetNodeID = strings.TrimSpace(targetNodeID)
targetIndex := -1
for index, node := range nodes {
if node.ID == targetNodeID && node.Type == design.NodeTypeImage && node.LayerRole == "image-generator" {
if node.ID == targetNodeID && node.Type == design.NodeTypeImage {
targetIndex = index
break
}
}
if targetIndex < 0 {
return append(nodes, generated...)
return append(nodes, completedImageGeneratorNodes(generated)...)
}
primary := generated[0]
if normalized, ok := completedImageGeneratorNode(primary); ok {
primary = normalized
}
target := nodes[targetIndex]
centerX := target.X + target.Width/2
centerY := target.Y + target.Height/2
@@ -4383,6 +4490,9 @@ func mergeGeneratedNodesIntoTarget(nodes []design.Node, generated []design.Node,
target.Tone = primary.Tone
target.Status = primary.Status
target.LayerRole = ""
if strings.TrimSpace(target.Tone) == "" || target.Tone == "image-generator" {
target.Tone = "visual"
}
if primary.Width > 0 {
target.Width = primary.Width
target.X = centerX - target.Width/2
@@ -4393,11 +4503,23 @@ func mergeGeneratedNodesIntoTarget(nodes []design.Node, generated []design.Node,
}
nodes[targetIndex] = target
if len(generated) > 1 {
return append(nodes, generated[1:]...)
return append(nodes, completedImageGeneratorNodes(generated[1:])...)
}
return nodes
}
func completedImageGeneratorNodes(nodes []design.Node) []design.Node {
next := make([]design.Node, len(nodes))
for index, node := range nodes {
if normalized, ok := completedImageGeneratorNode(node); ok {
next[index] = normalized
continue
}
next[index] = node
}
return next
}
func findNodeIndex(nodes []design.Node, nodeID string) int {
for index, node := range nodes {
if node.ID == nodeID {
@@ -554,6 +554,75 @@ func TestAgentChatRequiresTextPrompt(t *testing.T) {
}
}
func TestAgentChatImageModeBypassesPlannerAndHidesSettings(t *testing.T) {
ctx := context.Background()
repo := memory.NewProjectRepository()
planner := &recordingTaskPlanner{
fakeCreativeChatAgent: fakeCreativeChatAgent{
decision: design.AgentDecision{Intent: "image"},
},
}
service := NewDesignService(repo, nil, nil, nil, nil, planner)
queue := &fakeJobQueue{}
service.SetJobQueue(queue)
project := design.Project{
ID: "project-direct-image",
Title: "Direct image",
Status: design.StatusReady,
UpdatedAt: time.Now(),
}
if err := repo.Save(ctx, project); err != nil {
t.Fatal(err)
}
thread, err := service.AgentChat(ctx, project.ID, design.AgentChatRequest{
Mode: "image",
ImageSize: "1536x1024",
Messages: []design.AgentChatMessage{
{
Role: "user",
Contents: []design.AgentContent{
{Type: "text", Text: "a cute bird flatten ui", TextSource: "input"},
{Type: "text", Text: "图片生成参数:质量 自动,尺寸 1536x1024,比例 3:2,数量 3 张。纯生图模式,请直接围绕用户输入生成图片,不展开额外聊天。", TextSource: "settings"},
},
},
},
})
if err != nil {
t.Fatal(err)
}
if thread.Status != "running" {
t.Fatalf("expected running thread, got %s", thread.Status)
}
if planner.called != 0 {
t.Fatalf("expected image mode to bypass planner, got %d calls", planner.called)
}
if len(queue.jobs) != 1 {
t.Fatalf("expected one image job, got %#v", queue.jobs)
}
job := queue.jobs[0]
if job.TaskPlan.ImageCount != 3 || len(job.TaskPlan.ImageTasks) != 3 {
t.Fatalf("expected direct image count from settings, got %#v", job.TaskPlan)
}
if job.ImageSize != "1536x1024" {
t.Fatalf("expected image size to be preserved, got %s", job.ImageSize)
}
saved, err := repo.Get(ctx, project.ID)
if err != nil {
t.Fatal(err)
}
if len(saved.Messages) != 1 {
t.Fatalf("expected only the user message before generation starts, got %#v", saved.Messages)
}
if saved.Messages[0].Content != "a cute bird flatten ui" {
t.Fatalf("expected settings to stay out of visible messages, got %q", saved.Messages[0].Content)
}
if len(saved.Nodes) != 3 {
t.Fatalf("expected three direct image placeholders, got %#v", saved.Nodes)
}
}
func TestAgentChatCreatesPlaceholdersFromModelPlannedImageTasks(t *testing.T) {
ctx := context.Background()
repo := memory.NewProjectRepository()
@@ -50,6 +50,44 @@ func TestMergeGeneratedNodesIntoImageGeneratorTargetUsesGeneratedSizeAndKeepsCen
}
}
func TestMergeGeneratedNodesIntoCompletedImageGeneratorTargetDoesNotReappend(t *testing.T) {
nodes := []design.Node{
{
ID: "generator-1",
Type: design.NodeTypeImage,
Title: "旧图",
Content: "https://example.com/old.png",
Tone: "visual",
Status: "success",
LayerRole: "",
Width: 360,
Height: 240,
},
}
generated := []design.Node{
{
ID: "generator-1",
Type: design.NodeTypeImage,
Title: "新图",
Content: "https://example.com/new.png",
Tone: "image-generator",
Status: "success",
LayerRole: "image-generator",
Width: 1024,
Height: 1024,
},
}
merged := mergeGeneratedNodesIntoTarget(nodes, generated, "generator-1")
if len(merged) != 1 {
t.Fatalf("expected existing target to be updated in place, got %#v", merged)
}
node := merged[0]
if node.Content != "https://example.com/new.png" || node.LayerRole != "" || node.Tone != "visual" {
t.Fatalf("expected completed image generator to become regular image, got %#v", node)
}
}
func TestImageGeneratorTargetNodeIDParsesDirectiveCaseInsensitively(t *testing.T) {
targetID := imageGeneratorTargetNodeID([]design.AgentChatMessage{
{
@@ -64,6 +102,36 @@ func TestImageGeneratorTargetNodeIDParsesDirectiveCaseInsensitively(t *testing.T
}
}
func TestRestoreFinishedGeneratedImageNodesConvertsCompletedImageGeneratorNode(t *testing.T) {
now := time.Now()
project := design.Project{
ID: "project-completed-generator",
Status: design.StatusReady,
Nodes: []design.Node{
{
ID: "generator-1",
Type: design.NodeTypeImage,
Title: "生成图片",
Content: "http://localhost:19000/canvas-assets/cat.png",
Tone: "image-generator",
Status: "success",
LayerRole: "image-generator",
Width: 1024,
Height: 1024,
},
},
}
restored, changed := (*DesignService)(nil).restoreFinishedGeneratedImageNodes(project, now.Add(time.Second))
if !changed {
t.Fatal("expected completed image generator node to be normalized")
}
node := restored.Nodes[0]
if node.LayerRole != "" || node.Tone != "visual" || node.Status != "success" || node.Content != "http://localhost:19000/canvas-assets/cat.png" {
t.Fatalf("expected regular completed image node, got %#v", node)
}
}
func TestRestoreFinishedGeneratedImageNodesRepairsStaleCanvasSnapshot(t *testing.T) {
now := time.Now()
project := design.Project{