303 lines
10 KiB
TypeScript
303 lines
10 KiB
TypeScript
export type AgentPlanningBriefIntent =
|
||
| "image_edit"
|
||
| "vision"
|
||
| "image_gen"
|
||
| "video_gen"
|
||
| "chat";
|
||
|
||
export type AgentPlanningBriefAction =
|
||
| "respond"
|
||
| "clarify"
|
||
| "analyze_image"
|
||
| "generate_image"
|
||
| "edit_image"
|
||
| "generate_video";
|
||
|
||
export type AgentPlanningBriefTargetSource =
|
||
| "explicit_selection"
|
||
| "current_upload"
|
||
| "text_reference"
|
||
| "working_memory"
|
||
| "latest_result"
|
||
| "clarification"
|
||
| "none";
|
||
|
||
export type AgentPlanningBriefCapability =
|
||
| "respond"
|
||
| "analyze"
|
||
| "generic_edit"
|
||
| "remove_background"
|
||
| "remove_watermark"
|
||
| "enhance"
|
||
| "extend"
|
||
| "resize"
|
||
| "angle_variant"
|
||
| "local_edit"
|
||
| "box_cutout"
|
||
| "generate_image"
|
||
| "generate_video"
|
||
| "publish_canvas";
|
||
|
||
export interface AgentPlanningBriefInput {
|
||
userText: string;
|
||
intent: AgentPlanningBriefIntent;
|
||
subIntent?: string;
|
||
nextAction?: AgentPlanningBriefAction;
|
||
preferredCapability?: AgentPlanningBriefCapability;
|
||
targetSource?: AgentPlanningBriefTargetSource;
|
||
hasImage?: boolean;
|
||
imageFromHistory?: boolean;
|
||
selectedImageCount?: number;
|
||
currentUploadImageCount?: number;
|
||
historyImageCount?: number;
|
||
imageCount?: number;
|
||
aspectRatio?: string;
|
||
imageSize?: string;
|
||
planGoal?: string;
|
||
}
|
||
|
||
export interface AgentTaskGroup {
|
||
label: string;
|
||
count?: number;
|
||
description: string;
|
||
}
|
||
|
||
const CHINESE_NUMBER_VALUES: Record<string, number> = {
|
||
一: 1,
|
||
二: 2,
|
||
两: 2,
|
||
三: 3,
|
||
四: 4,
|
||
五: 5,
|
||
六: 6,
|
||
七: 7,
|
||
八: 8,
|
||
九: 9,
|
||
十: 10,
|
||
};
|
||
|
||
function compactLine(value: string, maxLength: number) {
|
||
const normalized = String(value || "").replace(/\s+/g, " ").trim();
|
||
if (normalized.length <= maxLength) return normalized;
|
||
return `${normalized.slice(0, maxLength)}...`;
|
||
}
|
||
|
||
function parseCountToken(token?: string | null) {
|
||
const value = String(token || "").trim();
|
||
if (!value) return undefined;
|
||
if (/^\d+$/.test(value)) {
|
||
const numeric = Number(value);
|
||
return Number.isFinite(numeric) && numeric > 0 ? Math.round(numeric) : undefined;
|
||
}
|
||
return CHINESE_NUMBER_VALUES[value];
|
||
}
|
||
|
||
function countFromText(text: string) {
|
||
const match = String(text || "").match(/([1-9]\d?|一|二|两|三|四|五|六|七|八|九|十)\s*张/);
|
||
return parseCountToken(match?.[1]);
|
||
}
|
||
|
||
function normalizeGroupDescription(description: string) {
|
||
return String(description || "")
|
||
.replace(/^[为是:：，,\s]+/, "")
|
||
.replace(/[。；;，,]\s*$/, "")
|
||
.trim();
|
||
}
|
||
|
||
export function extractAgentTaskGroups(text: string): AgentTaskGroup[] {
|
||
const normalized = String(text || "").replace(/\s+/g, " ").trim();
|
||
if (!normalized) return [];
|
||
|
||
const groups: AgentTaskGroup[] = [];
|
||
const groupPattern =
|
||
/第([一二两三四五六七八九十]|\d{1,2})组(?:为|是|:|：)?([\s\S]*?)(?=(?:第[一二两三四五六七八九十\d]{1,2}组)|[。；;]|$)/g;
|
||
let match: RegExpExecArray | null;
|
||
while ((match = groupPattern.exec(normalized)) !== null) {
|
||
const labelToken = match[1];
|
||
const description = normalizeGroupDescription(match[2] || "");
|
||
if (!description) continue;
|
||
const labelNumber = parseCountToken(labelToken);
|
||
groups.push({
|
||
label: `第${labelNumber || labelToken}组`,
|
||
count: countFromText(description),
|
||
description: compactLine(description, 96),
|
||
});
|
||
}
|
||
|
||
return groups.slice(0, 8);
|
||
}
|
||
|
||
function getTargetSourceLabel(source?: AgentPlanningBriefTargetSource) {
|
||
switch (source) {
|
||
case "explicit_selection":
|
||
return "画布选中图片";
|
||
case "current_upload":
|
||
return "当前上传图片";
|
||
case "text_reference":
|
||
return "文本指定图片";
|
||
case "working_memory":
|
||
return "会话工作记忆图片";
|
||
case "latest_result":
|
||
return "最近生成结果";
|
||
case "clarification":
|
||
return "待确认目标";
|
||
case "none":
|
||
default:
|
||
return "无需现有图片";
|
||
}
|
||
}
|
||
|
||
function getCapabilityLabel(capability?: AgentPlanningBriefCapability, subIntent?: string) {
|
||
switch (capability || subIntent) {
|
||
case "remove_background":
|
||
return "主体抠图/去背景";
|
||
case "remove_watermark":
|
||
return "去水印";
|
||
case "enhance":
|
||
return "清晰度增强";
|
||
case "extend":
|
||
return "扩图/画幅延展";
|
||
case "resize":
|
||
return "尺寸调整";
|
||
case "angle_variant":
|
||
case "generate_angle_variant":
|
||
return "多视角变体";
|
||
case "local_edit":
|
||
case "inpaint_region":
|
||
return "局部编辑";
|
||
case "box_cutout":
|
||
return "框选抠图";
|
||
case "generate_video":
|
||
return "视频生成";
|
||
case "generate_image":
|
||
return "图片生成";
|
||
case "analyze":
|
||
return "图片分析";
|
||
default:
|
||
return "图片编辑";
|
||
}
|
||
}
|
||
|
||
function inferTotalCount(params: AgentPlanningBriefInput, groups: AgentTaskGroup[]) {
|
||
if (typeof params.imageCount === "number" && Number.isFinite(params.imageCount)) {
|
||
return Math.max(1, Math.round(params.imageCount));
|
||
}
|
||
const groupedTotal = groups.reduce((sum, group) => sum + (group.count || 0), 0);
|
||
return groupedTotal > 0 ? groupedTotal : undefined;
|
||
}
|
||
|
||
function buildIntentLine(params: AgentPlanningBriefInput, groups: AgentTaskGroup[]) {
|
||
const target = getTargetSourceLabel(params.targetSource);
|
||
const capability = getCapabilityLabel(params.preferredCapability, params.subIntent);
|
||
const totalCount = inferTotalCount(params, groups);
|
||
const demand = compactLine(params.planGoal || params.userText || "", 120);
|
||
|
||
if (params.intent === "image_gen") {
|
||
const countText = totalCount ? `，共 ${totalCount} 张` : "";
|
||
const groupText = groups.length > 0 ? `，拆成 ${groups.length} 组` : "";
|
||
const referenceText = params.hasImage ? "，并保持参考图主体核心特征一致" : "";
|
||
return `用户需要生成新图片${groupText}${countText}${referenceText}；核心目标是：${demand || "完成本轮生图任务"}。`;
|
||
}
|
||
|
||
if (params.intent === "image_edit") {
|
||
return `用户需要基于${target}执行${capability}；核心目标是：${demand || "完成本轮图片处理"}。`;
|
||
}
|
||
|
||
if (params.intent === "video_gen") {
|
||
return `用户需要生成视频；先拆解镜头、主体连续性和运动节奏，再提交视频生成。`;
|
||
}
|
||
|
||
if (params.intent === "vision") {
|
||
return `用户需要理解或分析图片内容；先观察主体、场景、构图和潜在问题，再给出结论。`;
|
||
}
|
||
|
||
return `用户当前主要是在对话或咨询；先回答问题，不主动触发生图或修图工具。`;
|
||
}
|
||
|
||
function buildImageGenerationPlan(params: AgentPlanningBriefInput, groups: AgentTaskGroup[]) {
|
||
if (groups.length > 0) {
|
||
return groups.map((group, index) => {
|
||
const countText = group.count ? `${group.count}张` : "一组";
|
||
return `${index + 1}. ${group.label}：生成${countText}${group.description.replace(/^\d+\s*张/, "")}，保持主体识别一致。`;
|
||
});
|
||
}
|
||
|
||
const count = inferTotalCount(params, groups) || 1;
|
||
return [
|
||
`1. 锁定主体、场景、风格和构图约束，整理成可执行提示词。`,
|
||
`2. 生成 ${count} 张候选图，优先保证主体一致性和画面完成度。`,
|
||
`3. 输出结果并保留可继续追改的上下文。`,
|
||
];
|
||
}
|
||
|
||
function buildImageEditPlan(params: AgentPlanningBriefInput) {
|
||
const target = getTargetSourceLabel(params.targetSource);
|
||
const capability = getCapabilityLabel(params.preferredCapability, params.subIntent);
|
||
return [
|
||
`1. 先锁定${target}作为本轮处理目标，避免误处理其他图片。`,
|
||
`2. 调用${capability}能力执行核心改动，同时尽量保留主体结构和画质。`,
|
||
`3. 输出处理结果，并保留原图用于对比或继续微调。`,
|
||
];
|
||
}
|
||
|
||
function buildGenericPlan(params: AgentPlanningBriefInput) {
|
||
if (params.intent === "video_gen") {
|
||
return [
|
||
"1. 拆解视频主体、场景、动作节奏和镜头关系。",
|
||
"2. 生成脚本与分镜，并整理可直接提交的视频提示词。",
|
||
"3. 按规划提交视频生成，返回结果后支持继续调整。",
|
||
];
|
||
}
|
||
if (params.intent === "vision") {
|
||
return [
|
||
"1. 先识别图片主体、背景、构图和关键信息。",
|
||
"2. 结合用户问题输出判断、建议或可执行修改方向。",
|
||
];
|
||
}
|
||
return ["1. 先直接回答用户问题；如果需要执行工具，再根据明确指令进入下一步。"];
|
||
}
|
||
|
||
function buildOutputSettings(params: AgentPlanningBriefInput, groups: AgentTaskGroup[]) {
|
||
const lines: string[] = [];
|
||
const totalCount = inferTotalCount(params, groups);
|
||
if (params.intent === "image_gen" || params.intent === "image_edit") {
|
||
lines.push(`- 输出数量：${params.intent === "image_edit" ? "1 张处理结果" : `${totalCount || 1} 张`}`);
|
||
lines.push(`- 图片比例：${params.aspectRatio && params.aspectRatio !== "auto" ? params.aspectRatio : "按需求自适应"}`);
|
||
lines.push(`- 清晰度：${params.imageSize || (params.intent === "image_edit" ? "4K优先" : "2K优先")}`);
|
||
if (params.intent === "image_edit") {
|
||
lines.push(`- 目标来源：${getTargetSourceLabel(params.targetSource)}`);
|
||
}
|
||
} else if (params.intent === "video_gen") {
|
||
lines.push("- 输出形式：视频结果 + 可继续追改的分镜上下文");
|
||
} else {
|
||
lines.push("- 输出形式：直接回答，不主动触发生成工具");
|
||
}
|
||
return lines;
|
||
}
|
||
|
||
export function buildAgentPlanningBriefMarkdown(params: AgentPlanningBriefInput) {
|
||
const groups = extractAgentTaskGroups(params.userText);
|
||
const planLines =
|
||
params.intent === "image_gen"
|
||
? buildImageGenerationPlan(params, groups)
|
||
: params.intent === "image_edit"
|
||
? buildImageEditPlan(params)
|
||
: buildGenericPlan(params);
|
||
const outputSettings = buildOutputSettings(params, groups);
|
||
|
||
return [
|
||
"### 意图分析",
|
||
`- ${buildIntentLine(params, groups)}`,
|
||
"### 任务规划",
|
||
...planLines.map((line) => (line.startsWith("- ") ? line : `- ${line}`)),
|
||
"### 输出设置",
|
||
...outputSettings,
|
||
].join("\n");
|
||
}
|
||
|
||
export function isAgentPlanningBriefMarkdown(text: string) {
|
||
const normalized = String(text || "");
|
||
return /###\s*意图分析|【\s*意图分析\s*】/.test(normalized) &&
|
||
/###\s*任务规划|【\s*任务规划\s*】/.test(normalized);
|
||
}
|