docs: 更新 Agent 协议文档与部署配置

- 更新 Agent API 端点文档
- 更新 SSE 事件与输入输出文档
- 新增 deploy/.env.prod.example 配置模板
This commit is contained in:
qzl
2026-03-16 16:11:40 +08:00
parent 4b92772535
commit ed86bfe9ae
5 changed files with 384 additions and 978 deletions
+180 -337
View File
@@ -1,370 +1,213 @@
# Agent SSE Events
本文档记录 Agent Runtime 产生的所有 Server-Sent Events (SSE) 事件,用于前端实时展示
本文档描述 `GET /api/v1/agent/runs/{thread_id}/events` 的事件协议
## 事件流转架构
---
## 1) 事件管道
后端事件流转如下:
1. Runtime 直接产出 AG-UI 事件(如 `RUN_STARTED``TOOL_CALL_RESULT`
2. `agui_codec` 仅做协议对齐与字段净化(例如移除仅后端内部统计字段)
3. 事件同时:
- 持久化到数据库(用于 history)
- 发布到 Redis Stream(用于 SSE
4. `/runs/{thread_id}/events` 从 Redis Stream 读取并输出 SSE
---
## 2) SSE 帧格式
每条事件遵循标准 SSE
```text
id: <redis-stream-id>
event: <AG-UI-EVENT-TYPE>
data: <json>
```
pipeline.emit()
AgentScopeEventPipeline.emit()
├─→ store.persist() → 持久化到数据库
└─→ bus.publish() → 发布到 Redis Stream
前端通过 GET /runs/{thread_id}/events 读取
```
## 事件统一格式
- `id` 可用于断点续流(`Last-Event-ID`
- `event` 与 JSON 内 `type` 一致(例如 `RUN_STARTED`
- 空闲时可能出现 keep-alive 注释帧:
所有事件在 Redis 中传输时都包含以下字段:
```text
: keep-alive
```typescript
{
type: string, // 事件类型
threadId: string, // 会话 ID
runId: string, // 运行 ID
data: object // 事件数据
}
```
---
## 1. Orchestrator 生命周期事件
## 3) 事件类型(当前实现)
### run.started
### 3.1 Run 生命周期
Agent 开始运行时触发。
#### `RUN_STARTED`
```typescript
```json
{
type: "run.started",
threadId: "xxx",
runId: "yyy",
data: {}
"type": "RUN_STARTED",
"threadId": "...",
"runId": "..."
}
```
### run.finished
#### `RUN_FINISHED`
Agent 成功完成时触发。
```typescript
```json
{
type: "run.finished",
threadId: "xxx",
runId: "yyy",
data: {}
"type": "RUN_FINISHED",
"threadId": "...",
"runId": "..."
}
```
### run.error
#### `RUN_ERROR`
Agent 运行出错时触发。
```typescript
```json
{
type: "run.error",
threadId: "xxx",
runId: "yyy",
data: {
message: "runtime execution failed"
}
"type": "RUN_ERROR",
"threadId": "...",
"runId": "...",
"message": "runtime execution failed",
"code": null
}
```
### 3.2 阶段事件
#### `STEP_STARTED`
```json
{
"type": "STEP_STARTED",
"threadId": "...",
"runId": "...",
"stepName": "router" | "worker"
}
```
#### `STEP_FINISHED`
```json
{
"type": "STEP_FINISHED",
"threadId": "...",
"runId": "...",
"stepName": "router" | "worker"
}
```
### 3.3 Tool 事件
#### `TOOL_CALL_START`
```json
{
"type": "TOOL_CALL_START",
"threadId": "...",
"runId": "...",
"messageId": "...",
"toolCallId": "...",
"toolCallName": "...",
"stage": "worker"
}
```
#### `TOOL_CALL_ARGS`
```json
{
"type": "TOOL_CALL_ARGS",
"threadId": "...",
"runId": "...",
"messageId": "...",
"toolCallId": "...",
"toolCallName": "...",
"args": {},
"stage": "worker"
}
```
#### `TOOL_CALL_END`
```json
{
"type": "TOOL_CALL_END",
"threadId": "...",
"runId": "...",
"messageId": "...",
"toolCallId": "...",
"toolCallName": "...",
"stage": "worker"
}
```
#### `TOOL_CALL_RESULT`
```json
{
"type": "TOOL_CALL_RESULT",
"threadId": "...",
"runId": "...",
"messageId": "...",
"role": "tool",
"stage": "worker",
"tool_name": "...",
"tool_call_id": "...",
"tool_call_args": {},
"status": "success" | "failed",
"result_summary": "...",
"ui_schema": {},
"error": null
}
```
### 3.4 文本完成事件
#### `TEXT_MESSAGE_END`
当前实现仅在 worker 输出完成后发送完整结果,不发送 token delta 事件。
```json
{
"type": "TEXT_MESSAGE_END",
"threadId": "...",
"runId": "...",
"messageId": "...",
"role": "assistant",
"stage": "worker",
"status": "success" | "partial_success" | "failed",
"answer": "...",
"key_points": [],
"result_type": "execution_report" | "clarification" | "error_report" | "unknown",
"suggested_actions": [],
"error": null,
"ui_schema": {}
}
```
`inputTokens``outputTokens``cost``latencyMs``model` 属于后端内部统计字段,不在 SSE 对外协议中暴露。
### 3.5 快照事件
编码器支持以下 AG-UI 类型映射:
- `STATE_SNAPSHOT`
- `MESSAGES_SNAPSHOT`
当前 `/runs/{thread_id}/events` 主流程通常不主动产出这两类事件;历史查询请使用 `/history`
---
## 2. Step 阶段事件
## 4) 字段命名约定
### step.start
- 事件顶层通用字段使用 AG-UI 风格:`type``threadId``runId`
- 部分业务字段沿运行时模型历史命名保留下划线:
- `tool_name`
- `tool_call_id`
- `tool_call_args`
- `ui_schema`
阶段开始时触发
```typescript
{
type: "step.start",
threadId: "xxx",
runId: "yyy",
data: {
stepName: "router" | "worker"
}
}
```
### step.finish
阶段结束时触发。
```typescript
{
type: "step.finish",
threadId: "xxx",
runId: "yyy",
data: {
stepName: "router" | "worker"
}
}
```
---
## 3. Worker 运行时事件
### 3.1 文本消息事件
#### text.start
文本消息开始时触发。
```typescript
{
type: "text.start",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
role: "assistant",
stage: "worker"
}
}
```
#### text.delta
文本内容增量更新时触发。
```typescript
{
type: "text.delta",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
delta: "这是新增的文本内容",
stage: "worker"
}
}
```
#### text.end
文本消息结束时触发,包含完整输出和使用统计。
```typescript
{
type: "text.end",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
role: "assistant",
stage: "worker",
workerAgentOutput: {
status: "success" | "partial_success" | "failed",
answer: "主回复文本",
key_points: ["要点1", "要点2"],
result_type: "execution_report" | "clarification" | "error_report" | "unknown",
suggested_actions: ["建议操作1"],
error: null | { code: string, message: string },
ui_hints: { ... } | null // 仅在 ui_mode 非空时存在
},
model: "gpt-4o",
inputTokens: 1000,
outputTokens: 500,
cost: 0.025,
latencyMs: 1500
}
}
```
**workerAgentOutput 详细结构:**
```typescript
// WorkerAgentOutputLite
{
status: "success" | "partial_success" | "failed",
answer: string,
key_points: string[],
result_type: "execution_report" | "clarification" | "error_report" | "unknown",
suggested_actions: string[],
error: { code: string, message: string } | null
}
// WorkerAgentOutputRich (当 ui.ui_mode 非空时)
{
// ... WorkerAgentOutputLite 字段
ui_hints: {
ui_mode: string,
ui_state: object,
actions: UiHintAction[]
}
}
```
---
### 3.2 工具调用事件
#### tool.start
工具调用开始时触发。
```typescript
{
type: "tool.start",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
toolCallId: "call-abc",
toolName: "calendar_read",
stage: "worker"
}
}
```
#### tool.args
工具调用参数时触发。
```typescript
{
type: "tool.args",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
toolCallId: "call-abc",
toolName: "calendar_read",
args: { start_date: "2024-01-01", end_date: "2024-01-07" },
stage: "worker"
}
}
```
#### tool.end
工具调用结束时触发。
```typescript
{
type: "tool.end",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
toolCallId: "call-abc",
toolName: "calendar_read",
stage: "worker"
}
}
```
#### tool.result
工具执行结果时触发。
```typescript
{
type: "tool.result",
threadId: "xxx",
runId: "yyy",
data: {
messageId: "msg-xxx",
toolCallId: "call-abc",
toolName: "calendar_read",
stage: "worker",
toolAgentOutput: {
tool_name: "calendar_read",
tool_call_id: "call-abc",
tool_call_args: { start_date: "2024-01-01", end_date: "2024-01-07" },
status: "success" | "failed",
result_summary: "找到3个事件...",
ui_hints: null,
attachments: null
}
}
}
```
**toolAgentOutput 详细结构:**
```typescript
{
tool_name: string,
tool_call_id: string,
tool_call_args: object | null,
status: "success" | "failed",
result_summary: string,
ui_hints: object | null,
attachments: Array<{
bucket: string,
path: string,
mime_type: string,
url: string
}> | null
}
```
---
## 使用统计字段
`text.end` 事件中包含使用统计:
| 字段 | 类型 | 说明 |
|------|------|------|
| `model` | string | 使用的模型名称 |
| `inputTokens` | number | 输入 token 数量 |
| `outputTokens` | number | 输出 token 数量 |
| `cost` | number | 花费(美元) |
| `latencyMs` | number | 延迟(毫秒) |
---
## 前端接收示例
```javascript
const eventSource = new EventSource(`/runs/${threadId}/events`);
eventSource.addEventListener('run.started', (e) => {
console.log('Agent started:', JSON.parse(e.data));
});
eventSource.addEventListener('step.start', (e) => {
console.log('Step started:', JSON.parse(e.data));
});
eventSource.addEventListener('text.delta', (e) => {
const data = JSON.parse(e.data);
console.log('Text delta:', data.data.delta);
});
eventSource.addEventListener('tool.start', (e) => {
const data = JSON.parse(e.data);
console.log('Tool called:', data.data.toolName);
});
eventSource.addEventListener('tool.result', (e) => {
const data = JSON.parse(e.data);
console.log('Tool result:', data.data.toolAgentOutput);
});
eventSource.addEventListener('text.end', (e) => {
const data = JSON.parse(e.data);
console.log('Worker output:', data.data.workerAgentOutput);
console.log('Usage:', {
inputTokens: data.data.inputTokens,
outputTokens: data.data.outputTokens,
cost: data.data.cost
});
});
eventSource.addEventListener('run.finished', (e) => {
console.log('Agent finished:', JSON.parse(e.data));
});
eventSource.addEventListener('run.error', (e) => {
console.log('Agent error:', JSON.parse(e.data));
});
```
这部分命名属于当前后端实现约束,文档与实现保持一致