234 lines
9.0 KiB
TypeScript
234 lines
9.0 KiB
TypeScript
import { Socket } from "socket.io";
|
||
import { tool } from "ai";
|
||
import { z } from "zod";
|
||
import u from "@/utils";
|
||
import Memory from "@/utils/agent/memory";
|
||
import { createSkillTools, parseFrontmatter, scanSkills, useSkill } from "@/utils/agent/skillsTools";
|
||
import useTools from "@/agents/productionAgent/tools";
|
||
import ResTool from "@/socket/resTool";
|
||
import * as fs from "fs";
|
||
import path from "path";
|
||
|
||
export interface AgentContext {
|
||
socket: Socket;
|
||
isolationKey: string;
|
||
text: string;
|
||
userMessageTime?: number;
|
||
abortSignal?: AbortSignal;
|
||
resTool: ResTool;
|
||
msg: ReturnType<ResTool["newMessage"]>;
|
||
}
|
||
|
||
function buildMemPrompt(mem: Awaited<ReturnType<Memory["get"]>>): string {
|
||
let memoryContext = "";
|
||
if (mem.rag.length) {
|
||
memoryContext += `[相关记忆]\n${mem.rag.map((r) => r.content).join("\n")}`;
|
||
}
|
||
if (mem.summaries.length) {
|
||
if (memoryContext) memoryContext += "\n\n";
|
||
memoryContext += `[历史摘要]\n${mem.summaries.map((s, i) => `${i + 1}. ${s.content}`).join("\n")}`;
|
||
}
|
||
if (mem.shortTerm.length) {
|
||
if (memoryContext) memoryContext += "\n\n";
|
||
memoryContext += `[近期对话]\n${mem.shortTerm.map((m) => `${m.role}: ${m.content}`).join("\n")}`;
|
||
}
|
||
return `## Memory\n以下是你对用户的记忆，可作为参考但不要主动提及：\n${memoryContext}`;
|
||
}
|
||
|
||
export async function decisionAI(ctx: AgentContext) {
|
||
const { isolationKey, text, abortSignal } = ctx;
|
||
const memory = new Memory("productionAgent", isolationKey);
|
||
await memory.add("user", text);
|
||
|
||
const skill = path.join(u.getPath("skills"), "production_agent_decision.md");
|
||
const prompt = await fs.promises.readFile(skill, "utf-8");
|
||
|
||
const projectInfo = await u.db("o_project").where("id", ctx.resTool.data.projectId).first();
|
||
if (!projectInfo) throw new Error(`项目不存在，ID: ${ctx.resTool.data.projectId}`);
|
||
const [_, imageModelName] = projectInfo.imageModel!.split(":");
|
||
const [id, videoModelName] = projectInfo.videoModel!.split(":");
|
||
const data = await u.db("o_vendorConfig").where("id", id).select("models").first();
|
||
const models = JSON.parse(data!.models!);
|
||
const findData = models.find((i: any) => i.modelName == videoModelName);
|
||
const isRef = findData.mode.every((i: any) => Array.isArray(i));
|
||
const modelInfo = `项目使用的模型如下：\n图像模型：${imageModelName}\n视频模型：${videoModelName}\n多参：${isRef ? "是" : "否"}`;
|
||
|
||
const mem = buildMemPrompt(await memory.get(text));
|
||
|
||
const { textStream } = await u.Ai.Text("productionAgent").stream({
|
||
messages: [
|
||
{ role: "system", content: prompt },
|
||
{ role: "assistant", content: mem + "\n" + modelInfo },
|
||
{ role: "user", content: text },
|
||
],
|
||
abortSignal,
|
||
tools: {
|
||
...memory.getTools(),
|
||
...useTools({ resTool: ctx.resTool, msg: ctx.msg }),
|
||
...createSubAgent(ctx),
|
||
},
|
||
onFinish: async (completion) => {
|
||
await memory.add("assistant:decision", removeAllXmlTags(completion.text));
|
||
},
|
||
});
|
||
|
||
return textStream;
|
||
}
|
||
|
||
function createSubAgent(parentCtx: AgentContext) {
|
||
const { resTool, abortSignal } = parentCtx;
|
||
const memory = new Memory("productionAgent", parentCtx.isolationKey);
|
||
async function runAgent({
|
||
prompt,
|
||
system,
|
||
name,
|
||
memoryKey,
|
||
tools: extraTools,
|
||
messages,
|
||
}: {
|
||
prompt: string;
|
||
system: string;
|
||
name: string;
|
||
memoryKey: string;
|
||
tools?: Record<string, any>;
|
||
messages?: { role: "user" | "assistant" | "system"; content: string }[];
|
||
}) {
|
||
parentCtx.msg.complete();
|
||
const subMsg = resTool.newMessage("assistant", name);
|
||
const text = subMsg.text();
|
||
let fullResponse = "";
|
||
|
||
const { textStream } = await u.Ai.Text("scriptAgent").stream({
|
||
system,
|
||
messages: messages ?? [{ role: "user", content: prompt }],
|
||
abortSignal,
|
||
tools: { ...extraTools, ...useTools({ resTool, msg: subMsg }) },
|
||
});
|
||
|
||
try {
|
||
for await (const chunk of textStream) {
|
||
await new Promise<void>((resolve) => setTimeout(() => resolve(), 1));
|
||
text.append(chunk);
|
||
fullResponse += chunk;
|
||
}
|
||
text.complete();
|
||
subMsg.complete();
|
||
} catch (err: any) {
|
||
text.complete();
|
||
subMsg.stop();
|
||
throw err;
|
||
}
|
||
|
||
if (fullResponse.trim()) {
|
||
await memory.add(memoryKey, removeAllXmlTags(fullResponse), {
|
||
name,
|
||
createTime: new Date(subMsg.datetime).getTime(),
|
||
});
|
||
}
|
||
|
||
parentCtx.msg = resTool.newMessage("assistant", "视频策划");
|
||
return fullResponse;
|
||
}
|
||
|
||
const promptInput = z.object({
|
||
prompt: z.string().describe("交给子Agent的任务简约描述，100字以内"),
|
||
});
|
||
|
||
const run_sub_agent_execution = tool({
|
||
description: "执行层子Agent，负责衍生资产、",
|
||
inputSchema: promptInput,
|
||
execute: async ({ prompt }) => {
|
||
const skill = path.join(u.getPath("skills"), "production_agent_execution.md");
|
||
const systemPrompt = await fs.promises.readFile(skill, "utf-8");
|
||
const addPrompt =
|
||
"\n" +
|
||
[
|
||
"你必须使用如下XML格式写入工作区：\n```",
|
||
"拍摄计划：<scriptPlan>内容</scriptPlan>",
|
||
"分镜表：<storyboardTable>内容</storyboardTable>",
|
||
"分镜面板：<storyboardItem videoDesc='视频描述' prompt=提示词内容 track='分组' duration='视频推荐时间' associateAssetsIds='[该分镜所需的资产ID列表]'></storyboardItem>",
|
||
"```",
|
||
].join("\n");
|
||
|
||
const projectInfo = await u.db("o_project").where("id", resTool.data.projectId).first();
|
||
if (!projectInfo) throw new Error(`项目不存在，ID: ${resTool.data.projectId}`);
|
||
const artSkills = await createArtSkills(projectInfo?.artStyle!, projectInfo?.directorManual!);
|
||
|
||
const [_, imageModelName] = projectInfo.imageModel!.split(":");
|
||
const [id, videoModelName] = projectInfo.videoModel!.split(":");
|
||
const data = await u.db("o_vendorConfig").where("id", id).select("models").first();
|
||
const models = JSON.parse(data!.models!);
|
||
const findData = models.find((i: any) => i.modelName == videoModelName);
|
||
const isRef = findData.mode.every((i: any) => Array.isArray(i));
|
||
const modelInfo = `项目使用的模型如下：\n图像模型：${imageModelName}\n视频模型：${videoModelName}\n多参：${isRef ? "是" : "否"}`;
|
||
|
||
return runAgent({
|
||
prompt,
|
||
system: systemPrompt + addPrompt,
|
||
name: "执行导演",
|
||
memoryKey: "assistant:execution",
|
||
messages: [
|
||
{ role: "assistant", content: artSkills.prompt + `\n${modelInfo}` },
|
||
{ role: "user", content: prompt },
|
||
],
|
||
tools: { ...artSkills.tools },
|
||
});
|
||
},
|
||
});
|
||
|
||
const run_sub_agent_supervision = tool({
|
||
description: "监制层子Agent，负责审核执行结果",
|
||
inputSchema: promptInput,
|
||
execute: async ({ prompt }) => {
|
||
const skill = path.join(u.getPath("skills"), "production_agent_supervision.md");
|
||
const systemPrompt = await fs.promises.readFile(skill, "utf-8");
|
||
return runAgent({
|
||
prompt,
|
||
system: systemPrompt + "你必须使用如下XML格式写入工作区：\n<storySkeleton>故事骨架内容</storySkeleton>",
|
||
name: "监制",
|
||
memoryKey: "assistant:supervision",
|
||
});
|
||
},
|
||
});
|
||
|
||
return { run_sub_agent_execution, run_sub_agent_supervision };
|
||
}
|
||
|
||
async function createArtSkills(artName: string, storyName: string) {
|
||
const artWorkerPath = u.getPath(["skills", "art_skills", artName, "driector_skills"]);
|
||
const storyWorkerPath = u.getPath(["skills", "story_skills", storyName, "driector_skills"]);
|
||
const skillList = [...(await scanSkills(artWorkerPath + "/*.md")), ...(await scanSkills(storyWorkerPath + "/*.md"))];
|
||
const mainSkills: { path: string; name: string; description: string }[] = [];
|
||
for (const skillPath of skillList) {
|
||
if (!fs.existsSync(skillPath)) throw new Error(`主技能文件不存在: ${skillPath}`);
|
||
const content = await fs.promises.readFile(skillPath, "utf-8");
|
||
const parsed = parseFrontmatter(content);
|
||
mainSkills.push({ path: skillPath, ...parsed });
|
||
}
|
||
const res = {
|
||
prompt: `## Skills
|
||
以下技能提供了专业任务的专用指令。
|
||
当任务与某个技能的描述匹配时，调用 activate_skill 工具并传入技能名称来加载完整指令。
|
||
加载后遵循技能指令执行任务，需要时调用 read_skill_file 读取资源文件内容。
|
||
${buildSkillPrompt(mainSkills)}`,
|
||
tools: createSkillTools(mainSkills, { mainSkill: mainSkills, secondarySkills: [], tertiarySkills: [] }),
|
||
};
|
||
return res;
|
||
}
|
||
|
||
function removeAllXmlTags(text: string): string {
|
||
text = text.replace(/<([a-zA-Z][\w-]*)(\s+[^>]*)?>([\s\S]*?)<\/\1>/g, "");
|
||
text = text.replace(/<([a-zA-Z][\w-]*)(\s+[^>]*)?\/>/g, "");
|
||
text = text.replace(/<\/?[a-zA-Z][\w-]*(\s+[^>]*)?>/g, "");
|
||
return text.trim();
|
||
}
|
||
|
||
export function buildSkillPrompt(skills: { name: string; description: string }[]): string {
|
||
const skillEntries = skills
|
||
.map((s) => ` <skill>\n <name>${s.name}</name>\n <description>${s.description}</description>\n </skill>`)
|
||
.join("\n");
|
||
return `
|
||
<available_skills>
|
||
${skillEntries}
|
||
</available_skills>`;
|
||
} |