feat: 实现 AgentScope ReAct Runner 两阶段执行并重构事件处理

This commit is contained in:
zl-q
2026-03-16 09:01:01 +08:00
parent 072c09d99d
commit dcceb48d84
51 changed files with 5015 additions and 5663 deletions
+432
View File
@@ -0,0 +1,432 @@
# 前后端数据流通指南
本文档描述前端如何与后端 Agent 系统交互,以及如何渲染后端返回的 UI 数据。
---
## 1. 整体交互流程
```
┌─────────────────────────────────────────────────────────────────────────┐
│ 前端 │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────┐ POST /runs ┌──────────────┐ │
│ │ 构造请求 │ ──────────────────▶│ 后端 │ │
│ │ (多模态数据) │ │ (任务入队) │ │
│ └──────────────┘◀────────────────── └──────────────┘ 202 Accepted │
│ │ taskId │
│ │ │
│ │ GET /runs/{threadId}/events │
│ │──────────────────────────────────────────────────────────▶ │
│ │ SSE 事件流 ◀───────────────── │
│ │ │
│ ┌─────┴──────────────────────────────────────────────────────────┐ │
│ │ 渲染事件 │ │
│ │ - text.delta: 追加文本 │ │
│ │ - tool.result: 渲染 toolAgentOutput.ui_hints → UiSchema │ │
│ │ - text.end: 渲染 workerAgentOutput.ui_hints → UiSchema │ │
│ └────────────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────────────┐
│ 后端 │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ POST /runs │
│ → 验证请求 │
│ → 任务入队列 (Taskiq) │
│ │
│ Worker 执行 │
│ → emit 事件到 Redis Stream │
│ → UiHints → ui_compiler.compile() → UiSchemaRenderer │
│ │
│ 持久化 │
│ → AgentChatMessage (含 uiSchema) │
│ │
└─────────────────────────────────────────────────────────────────────────┘
```
---
## 2. 多模态数据处理
### 2.1 前端上传附件流程
```javascript
// 方式一:先上传附件,获取签名 URL
const formData = new FormData();
formData.append('threadId', threadId);
formData.append('file', imageFile);
const uploadResponse = await fetch('/api/v1/agent/attachments', {
method: 'POST',
body: formData,
});
const { attachment } = await uploadResponse.json();
// attachment.url 是临时访问 URL
// 方式二:直接使用已有的签名 URL(如果有)
const imageUrl = 'https://storage.example.com/...';
```
### 2.2 构造 RunAgentInput
```javascript
const runInput = {
threadId: threadId,
runId: `run-${Date.now()}`,
state: {},
messages: [
{
id: `msg-${Date.now()}`,
role: 'user',
content: [
{ type: 'text', text: '这张图片里有什么?' },
{
type: 'binary',
mimeType: 'image/png',
url: attachment.url // 签名 URL
}
]
}
],
tools: [], // 或传入工具定义
context: [],
forwardedProps: {}
};
```
### 2.3 发送请求
```javascript
const response = await fetch('/api/v1/agent/runs', {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify(runInput)
});
// 返回 202 Accepted
const { taskId, threadId, runId, created } = await response.json();
```
---
## 3. 事件流渲染
### 3.1 订阅事件
```javascript
class AgentEventHandler {
constructor(threadId) {
this.eventSource = new EventSource(`/api/v1/agent/runs/${threadId}/events`);
this.setupListeners();
}
setupListeners() {
this.eventSource.addEventListener('run.started', this.onRunStarted);
this.eventSource.addEventListener('step.start', this.onStepStart);
this.eventSource.addEventListener('text.delta', this.onTextDelta);
this.eventSource.addEventListener('text.message.start', this.onTextStart);
this.eventSource.addEventListener('tool.call.start', this.onToolStart);
this.eventSource.addEventListener('tool.call.args', this.onToolArgs);
this.eventSource.addEventListener('tool.call.result', this.onToolResult);
this.eventSource.addEventListener('text.message.end', this.onTextEnd);
this.eventSource.addEventListener('run.finished', this.onRunFinished);
this.eventSource.addEventListener('run.error', this.onRunError);
}
}
```
### 3.2 事件处理与渲染
#### text.delta - 文本增量
```javascript
onTextDelta(event) {
const data = JSON.parse(event.data);
const { delta, messageId } = data.data;
// 追加到对应消息的文本内容
appendTextToMessage(messageId, delta);
}
```
#### tool.result - 工具结果
```javascript
onToolResult(event) {
const data = JSON.parse(event.data);
const { toolAgentOutput, toolCallId } = data.data;
if (toolAgentOutput?.ui_hints) {
// 后端返回的是 UiHints,需要编译为 UiSchemaRenderer
const uiSchema = compileUiHints(toolAgentOutput.ui_hints);
renderUiSchema(uiSchema, toolCallId);
} else {
// 纯文本结果
renderText(toolAgentOutput.result_summary, toolCallId);
}
}
```
#### text.end - Worker 完成
```javascript
onTextEnd(event) {
const data = JSON.parse(event.data);
const { workerAgentOutput, inputTokens, outputTokens, cost } = data.data;
if (workerAgentOutput?.ui_hints) {
// 渲染 UiHints → UiSchemaRenderer
const uiSchema = compileUiHints(workerAgentOutput.ui_hints);
renderUiSchema(uiSchema, messageId);
} else {
// 纯文本回复
renderText(workerAgentOutput?.answer, messageId);
}
// 显示使用统计
showUsageStats({ inputTokens, outputTokens, cost });
}
```
---
## 4. History 对话历史渲染
### 4.1 获取历史
```javascript
const response = await fetch(
`/api/v1/agent/history?threadId=${threadId}&before=${date}`
);
const { messages } = await response.json();
```
### 4.2 渲染历史消息
```javascript
messages.forEach(msg => {
switch (msg.role) {
case 'user':
renderUserMessage(msg);
break;
case 'assistant':
if (msg.uiSchema) {
// 直接渲染 UiSchemaRenderer(后端已编译好)
renderUiSchema(msg.uiSchema);
} else {
// 纯文本
renderText(msg.content);
}
break;
case 'tool':
// tool 结果也可能有 uiSchema
if (msg.uiSchema) {
renderUiSchema(msg.uiSchema);
} else {
renderText(msg.content);
}
break;
}
});
```
---
## 5. UiSchemaRenderer 渲染
### 5.1 数据结构
后端 `ui_compiler.compile()` 输出的结构:
```typescript
interface UiSchemaRenderer {
version: "2.0";
locale: string;
status: "info" | "success" | "warning" | "error" | "pending";
theme: "default" | "light" | "dark";
meta?: {
requestId?: string;
toolId?: string;
traceId?: string;
userId?: string;
};
root: UiLayoutNode;
}
// 布局节点
interface UiStackNode {
type: "stack";
direction: "vertical" | "horizontal";
gap?: number;
appearance: "plain" | "card" | "section";
status?: UiStatus;
align?: LayoutAlign;
justify?: LayoutJustify;
wrap?: boolean;
children: UiNode[];
}
interface UiGridNode {
type: "grid";
columns: number;
gap?: number;
appearance: LayoutAppearance;
status?: UiStatus;
children: UiNode[];
}
// 基础节点
type UiNode =
| UiTextNode // 文本
| UiIconNode // 图标
| UiBadgeNode // 标签
| UiButtonNode // 按钮
| UiKvNode // 键值对
| UiDividerNode; // 分割线
// 文本节点
interface UiTextNode {
type: "text";
content: string;
format: "plain" | "markdown";
role: "title" | "subtitle" | "body" | "caption" | "code";
status?: UiStatus;
maxLines?: number;
visible?: boolean;
}
// 按钮节点
interface UiButtonNode {
type: "button";
label: string;
style: "primary" | "secondary" | "ghost" | "danger";
disabled?: boolean;
icon?: UiIconSpec;
action: UiActionPayload;
}
// 键值对节点
interface UiKvNode {
type: "kv";
items: UiKvItem[];
columns?: number;
}
interface UiKvItem {
key: string;
label?: string;
value: any;
copyable?: boolean;
}
// Action payloads
type UiActionPayload =
| { type: "navigation"; path: string; params?: Record<string, any> }
| { type: "url"; url: string; target?: "_self" | "_blank" }
| { type: "event"; event: string; payload?: Record<string, any> }
| { type: "tool"; toolId: string; params?: Record<string, any> }
| { type: "copy"; content: string; successMessage?: string }
| { type: "payload"; payload: Record<string, any>; submitTo?: string };
```
### 5.2 统一渲染器实现
```dart
class UiSchemaRendererWidget extends StatelessWidget {
final UiSchemaRenderer schema;
@override
Widget build(BuildContext context) {
return renderNode(schema.root);
}
Widget renderNode(UiNode node) {
switch (node.type) {
case 'text':
return renderTextNode(node as UiTextNode);
case 'icon':
return renderIconNode(node as UiIconNode);
case 'badge':
return renderBadgeNode(node as UiBadgeNode);
case 'button':
return renderButtonNode(node as UiButtonNode);
case 'kv':
return renderKvNode(node as UiKvNode);
case 'divider':
return renderDividerNode(node as UiDividerNode);
case 'stack':
return renderStackNode(node as UiStackNode);
case 'grid':
return renderGridNode(node as UiGridNode);
default:
return SizedBox.shrink();
}
}
}
```
### 5.3 事件渲染 vs History 渲染一致性
**关键点**:两者使用同一个渲染器
| 数据来源 | 数据格式 | 处理方式 |
|---------|---------|---------|
| events text.end | `workerAgentOutput.ui_hints` | compile → UiSchemaRenderer → 渲染 |
| events tool.result | `toolAgentOutput.ui_hints` | compile → UiSchemaRenderer → 渲染 |
| history assistant | `msg.uiSchema` | 直接渲染 |
| history tool | `msg.uiSchema` | 直接渲染 |
**编译函数**(前端需要实现):
```javascript
// 将后端 UiHints 编译为 UiSchemaRenderer
function compileUiHints(hints) {
// 这里可以调用后端的 compile 接口
// 或者在前端实现相同的编译逻辑
return fetch('/api/v1/agent/compile-ui-hints', {
method: 'POST',
body: JSON.stringify(hints)
}).then(r => r.json());
}
```
---
## 6. UiHints 介绍(可选)
如果前端需要自己处理 UiHints,以下是结构:
### UiHintsPayload
```typescript
interface UiHintsPayload {
version: "2.1";
intent: "message" | "data" | "list" | "status" | "form" | "mixed";
status: "info" | "success" | "warning" | "error" | "pending";
title?: string;
description?: string;
body?: string;
bodyFormat?: "plain" | "markdown";
items?: UiHintKvItem[];
listItems?: UiHintListItem[];
sections?: UiHintSection[];
actions?: UiHintAction[];
icon?: UiHintIcon;
meta?: Record<string, any>;
}
```
**注意**History API 返回的消息已经包含了编译好的 `uiSchema`,前端不需要自己编译。
---
## 7. 总结
1. **多模态**:前端上传附件 → 获取签名 URL → 构造 RunAgentInput → POST /runs
2. **实时事件**:订阅 /runs/{threadId}/events → 处理 text.delta/tool.result/text.end → 渲染 UiSchema
3. **历史消息**GET /history → 遍历消息 → 有 uiSchema 则渲染,否则渲染纯文本
4. **一致性**events 和 history 使用同一个 UiSchemaRenderer 渲染组件