mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-16 21:22:01 +00:00
协作团队真的能并行了,界面上也看得出来
先说一个和名字不符的事实:这个节点叫「多 Agent 协作」,但在此之前它是一条
严格的串行流水线。调度协议写死成 `next: string`——"一次只能派一个"被编进了
协议本身,模型再想并行也表达不出来。三个专家各花 10 秒,总是 30 秒。
(真并行一直在图级 fan-out:实测 3 个各 sleep 2 秒的节点连在同一个上游,
全部 +0.00s 开始、+2.37s 结束,墙钟 2.6 秒。只是协作团队走不到那条路。)
**协议改成 assignments 数组**,一轮可以派多个互不依赖的专家,asyncio.gather
并发执行,总耗时按最慢的那个算。max_parallel 可配,设成 1 即退回严格串行。
**并发只在任务真的互不依赖时才成立**,这件事的成本必须说清楚:同一轮发出去的
专家拿到的是**同一份进展快照**,谁也看不见谁这一轮干了什么。派错了不报错,
只会让两个人基于一样的旧信息重复劳动,而这在结果里很难看出来。所以判断责任
压在调度者身上,提示词里把反例写出来:
可以同时派:「查 A 的资料」和「查 B 的资料」
不能同时派:「查资料」和「根据资料写报告」
不能同时派:「写初稿」和「校对初稿」
并明确告诉它拿不准就分两轮——代价只是慢一点。
**每个人单独计时**,而不是事后拿整轮墙钟去摊。界面要显示"并行省下多少",
那个数必须是量出来的:按 N×墙钟 推算等于假设每个人都跑满了最慢那条,
而实际上快的那个可能只花了三分之一。测试里专门有个快慢不对称的用例守这条。
**调度日志改成结构化字段**(agents / done / reason / parallel)。以前只有一句
中文消息串,解码层得拿正则去拆「调度 → X(理由)」——后端文案一改就散架,
而这次正是改了文案。老运行的事件没有这些字段,正则那条路留着兜底。
**界面加泳道。** 每个成员一行、每轮一格,条形宽度按耗时归一化,所以一眼看得出
谁和谁是同时干的、谁是这一轮的瓶颈。顶上一行汇总"3 名成员 · 3 轮 · 2 轮并行,
省下 11.8s";点任意一格展开那个人这一轮的任务和回复。全程串行时不装作有并行,
而是说"本次任务互相依赖,逐轮进行"。
泳道排在子步骤**前面**:先给一眼看清的形状,再往下是每一步的内容。反过来的话,
得把十几行步骤读完才知道这个团队是怎么分工的。窄栏(画布右栏 360px)下隐去
"N 人同时"的文字——并行与否看条形本来就一目了然,而缩写成"并2"没人看得懂。
验证:后端 276 passed(新增 test_supervisor.py 7 条,核心那条用**墙钟**判定
并发而不是看事件顺序——顺序只能说明"发出去了",说明不了"同时跑");
check-stream 新增 10 条泳道断言,含"同一轮里慢的那个条更长"和两种栏宽不溢出;
tsc 干净,三个 check 不回归。
This commit is contained in:
@@ -161,10 +161,16 @@ export const NODE_DEFS: Record<NodeType, NodeDef> = {
|
||||
{ key: 'goal', label: '团队目标', type: 'prompt' },
|
||||
{ key: 'agents', label: '团队成员', type: 'agents' },
|
||||
{ key: 'max_rounds', label: '最大轮次', type: 'number', min: 1, max: 25 },
|
||||
{
|
||||
key: 'max_parallel', label: '每轮最多同时派几人', type: 'number', min: 1, max: 6,
|
||||
help: '互不依赖的任务调度者可以放在同一轮同时执行,总耗时按最慢的那个算。'
|
||||
+ '设为 1 即退回严格串行——并发的成员看到的是同一份进展快照,'
|
||||
+ '任务其实有依赖却被同时派出去,两个人会基于一样的旧信息重复劳动',
|
||||
},
|
||||
...MODEL_FIELDS,
|
||||
...COMMON_TAIL,
|
||||
],
|
||||
defaults: { max_rounds: 6, agents: [] },
|
||||
defaults: { max_rounds: 6, max_parallel: 3, agents: [] },
|
||||
},
|
||||
tool: {
|
||||
type: 'tool', label: '调用工具', category: '执行', icon: Wrench,
|
||||
|
||||
@@ -136,6 +136,50 @@ const REVIEW_DEGRADED_TURN: StreamTurn = {
|
||||
rawOutput: { answer: "平台管理员可以修改他人权限。" },
|
||||
}
|
||||
|
||||
/** 协作团队的泳道:前两轮并行、第三轮汇总 */
|
||||
const TEAM_TURN: StreamTurn = {
|
||||
id: "team", question: "比较三种向量数据库在中文检索上的取舍", phase: "done", status: "完成",
|
||||
steps: [{
|
||||
id: "n-team", seq: 1, kind: "node", status: "done", title: "调研团队", meta: "24.1s",
|
||||
team: {
|
||||
members: ["researcher", "analyst", "writer"],
|
||||
savedMs: 11800,
|
||||
finished: true,
|
||||
rounds: [
|
||||
{
|
||||
round: 0, parallel: 2, wallMs: 9600, sumMs: 16400,
|
||||
reason: "两边资料互不相干,可以同时查",
|
||||
members: [
|
||||
{ agent: "researcher", instruction: "查 Milvus 与 Qdrant 的中文分词支持", ms: 9600,
|
||||
status: "done", result: "Milvus 2.4 起内置 jieba…" },
|
||||
{ agent: "analyst", instruction: "查 pgvector 的中文检索方案", ms: 6800,
|
||||
status: "done", result: "pgvector 本身不分词,需配 zhparser…" },
|
||||
],
|
||||
},
|
||||
{
|
||||
round: 1, parallel: 2, wallMs: 7400, sumMs: 12400,
|
||||
reason: "两组基准测试互不依赖",
|
||||
members: [
|
||||
{ agent: "researcher", instruction: "跑召回率基准", ms: 7400, status: "done",
|
||||
result: "hit@1:Milvus 0.91 / Qdrant 0.89 / pgvector 0.84" },
|
||||
{ agent: "analyst", instruction: "跑写入吞吐基准", ms: 5000, status: "done",
|
||||
result: "10 万条:Milvus 42s / Qdrant 51s / pgvector 88s" },
|
||||
],
|
||||
},
|
||||
{
|
||||
round: 2, parallel: 1, wallMs: 7100, sumMs: 7100,
|
||||
reason: "两份基准都出来了,可以汇总",
|
||||
members: [
|
||||
{ agent: "writer", instruction: "根据前面的结论写选型建议", ms: 7100,
|
||||
status: "done", result: "## 结论\n中文为主且要开箱即用 → Milvus…" },
|
||||
],
|
||||
},
|
||||
],
|
||||
},
|
||||
}],
|
||||
output: { 建议: "中文为主且要开箱即用选 Milvus;已有 Postgres 且数据量在百万以内,pgvector + zhparser 的运维成本最低。" },
|
||||
}
|
||||
|
||||
const LONG_TURN: StreamTurn = {
|
||||
id: "long", question: "需要能够显示总结内容", phase: "done", status: "完成",
|
||||
steps: [], output: { result: longReport() },
|
||||
@@ -185,6 +229,9 @@ function Preview() {
|
||||
if (params.get('review') === '1') {
|
||||
return <AssistantStream turns={[REVIEW_BROKEN_TURN, REVIEW_DEGRADED_TURN]} />
|
||||
}
|
||||
if (params.get('team') === '1') {
|
||||
return <AssistantStream turns={[TEAM_TURN]} dense={params.get('dense') === '1'} />
|
||||
}
|
||||
|
||||
return (
|
||||
<div className="flex h-full flex-col">
|
||||
|
||||
@@ -1,12 +1,13 @@
|
||||
import { useEffect, useRef, useState, type ReactNode } from 'react'
|
||||
import {
|
||||
AlertTriangle, Brain, ChevronRight, CircleCheck, CircleDot, Database,
|
||||
FileCode, GitBranch, Hand, Info, Sparkles, Table2, Terminal, Wrench, XCircle,
|
||||
FileCode, GitBranch, Hand, Info, Sparkles, Table2, Terminal, Users, Wrench, XCircle,
|
||||
} from 'lucide-react'
|
||||
import clsx from 'clsx'
|
||||
import { Spinner } from '../components/ui'
|
||||
import {
|
||||
parseQueryResult, type ResultTable as Table, type Step, type StepKind,
|
||||
type TeamRun,
|
||||
} from './decode'
|
||||
import { Markdown } from './Markdown'
|
||||
import type { ReviewResult } from '../types'
|
||||
@@ -317,6 +318,11 @@ function StepRow({ step, dense, depth }: { step: Step; dense: boolean; depth: nu
|
||||
</div>
|
||||
)}
|
||||
|
||||
{/* 泳道排在子步骤**前面**:先给一眼看清的形状(谁做了哪几轮、哪些是
|
||||
同时进行的、各花多久),再往下是每一步的内容。反过来的话,得把
|
||||
十几行步骤读完才知道这个团队是怎么分工的 */}
|
||||
{step.team && <TeamLanes team={step.team} dense={dense} />}
|
||||
|
||||
{!!step.children?.length && (
|
||||
<StepList steps={step.children} dense={dense} depth={depth + 1} />
|
||||
)}
|
||||
@@ -324,6 +330,129 @@ function StepRow({ step, dense, depth }: { step: Step; dense: boolean; depth: nu
|
||||
)
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* 协作团队的泳道:每个成员一行,每一轮一格。
|
||||
*
|
||||
* 这个节点以前在界面上是一条扁平的步骤序列——能看出"谁回了什么",看不出
|
||||
* "谁和谁是同时干的"。而一轮同时派几个人正是它相对单 agent 的全部优势,
|
||||
* 不画出来等于没有。
|
||||
*
|
||||
* 条形的宽度按耗时归一化,所以一眼能看出谁是这一轮的瓶颈。
|
||||
*/
|
||||
function TeamLanes({ team, dense }: { team: TeamRun; dense: boolean }) {
|
||||
const [openKey, setOpenKey] = useState<string | null>(null)
|
||||
if (!team.rounds.length) return null
|
||||
|
||||
// 所有格子共用一个时间标尺,否则"快的那格"和"慢的那格"画得一样长,
|
||||
// 瓶颈就看不出来了
|
||||
const slowest = Math.max(1, ...team.rounds.flatMap((r) => r.members.map((m) => m.ms)))
|
||||
const parallelRounds = team.rounds.filter((r) => r.parallel > 1).length
|
||||
|
||||
return (
|
||||
<div className="mt-2 rounded-lg border bg-bg p-2">
|
||||
<div className="mb-2 flex flex-wrap items-center gap-x-2.5 gap-y-1 text-[10.5px] text-faint">
|
||||
<span className="flex items-center gap-1 text-dim">
|
||||
<Users size={11} /> {team.members.length} 名成员 · {team.rounds.length} 轮
|
||||
</span>
|
||||
{parallelRounds > 0 && (
|
||||
<span style={{ color: 'var(--ok)' }}>
|
||||
{parallelRounds} 轮并行,省下 {fmtMs(team.savedMs)}
|
||||
</span>
|
||||
)}
|
||||
{parallelRounds === 0 && team.rounds.length > 1 && (
|
||||
// 全程串行不是故障,但值得说一句——任务本来就互相依赖时它就该是串行的
|
||||
<span>本次任务互相依赖,逐轮进行</span>
|
||||
)}
|
||||
</div>
|
||||
|
||||
<div className="space-y-1">
|
||||
{team.members.map((name) => (
|
||||
<div key={name} className="flex items-center gap-1.5">
|
||||
<div className={clsx('shrink-0 truncate text-[10.5px] text-dim',
|
||||
dense ? 'w-14' : 'w-20')}
|
||||
title={name}>
|
||||
{name}
|
||||
</div>
|
||||
<div className="flex min-w-0 flex-1 gap-1">
|
||||
{team.rounds.map((r) => {
|
||||
const m = r.members.find((x) => x.agent === name)
|
||||
const key = `${r.round}:${name}`
|
||||
return (
|
||||
<div key={r.round} className="min-w-0 flex-1">
|
||||
{m ? (
|
||||
<button
|
||||
className="h-4 w-full rounded-sm text-left transition-opacity hover:opacity-80"
|
||||
title={`第 ${r.round + 1} 轮 · ${m.instruction || '(无指令)'}`}
|
||||
onClick={() => setOpenKey(openKey === key ? null : key)}
|
||||
style={{
|
||||
background: m.status === 'running'
|
||||
? 'color-mix(in srgb, var(--accent) 45%, transparent)'
|
||||
: 'var(--accent)',
|
||||
// 宽度按耗时占比,但留一个下限,否则快的那格细到看不见
|
||||
width: `${Math.max(18, (m.ms / slowest) * 100)}%`,
|
||||
}}
|
||||
/>
|
||||
) : (
|
||||
<div className="h-4 w-full rounded-sm"
|
||||
style={{ background: 'var(--hover)' }} />
|
||||
)}
|
||||
</div>
|
||||
)
|
||||
})}
|
||||
</div>
|
||||
</div>
|
||||
))}
|
||||
|
||||
<div className="flex items-center gap-1.5 pt-0.5">
|
||||
<div className={clsx('shrink-0', dense ? 'w-14' : 'w-20')} />
|
||||
<div className="flex min-w-0 flex-1 gap-1">
|
||||
{team.rounds.map((r) => (
|
||||
<div key={r.round}
|
||||
className="min-w-0 flex-1 truncate text-center text-[9.5px] text-faint"
|
||||
title={r.reason || ''}>
|
||||
{/* 窄栏放不下完整措辞,但"并2"这种缩写没人看得懂,
|
||||
宁可只留轮次和耗时,并行与否看上面那行条形本来就一目了然 */}
|
||||
{`第${r.round + 1}轮`}
|
||||
{r.parallel > 1 && !dense ? ` · ${r.parallel} 人同时` : ''}
|
||||
{' · '}{fmtMs(r.wallMs)}
|
||||
</div>
|
||||
))}
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
{openKey && (() => {
|
||||
const [rd, name] = openKey.split(':')
|
||||
const m = team.rounds.find((r) => String(r.round) === rd)
|
||||
?.members.find((x) => x.agent === name)
|
||||
if (!m) return null
|
||||
return (
|
||||
<div className="mt-2 rounded border bg-panel p-2 text-[10.5px] leading-relaxed">
|
||||
<div className="mb-1 text-dim">
|
||||
{name} · 第 {Number(rd) + 1} 轮 · {fmtMs(m.ms)}
|
||||
</div>
|
||||
{m.instruction && (
|
||||
<div className="mb-1.5 text-faint">任务:{m.instruction}</div>
|
||||
)}
|
||||
{m.result && (
|
||||
<pre className="mono max-h-36 overflow-auto whitespace-pre-wrap text-dim">
|
||||
{m.result.slice(0, 1200)}
|
||||
</pre>
|
||||
)}
|
||||
</div>
|
||||
)
|
||||
})()}
|
||||
</div>
|
||||
)
|
||||
}
|
||||
|
||||
/** 毫秒 → 人读的单位。和 decode 里的 dur 同一套规则 */
|
||||
function fmtMs(ms: number): string {
|
||||
if (ms < 1000) return `${ms}ms`
|
||||
return ms < 60_000 ? `${(ms / 1000).toFixed(1)}s` : `${Math.round(ms / 60_000)}min`
|
||||
}
|
||||
|
||||
/** 查询结果画成表格。一屏 JSON 谁也看不出名堂,表格能一眼看到形状。 */
|
||||
function ResultTable({ table, dense }: { table: Table; dense: boolean }) {
|
||||
const maxCols = dense ? 4 : 8
|
||||
|
||||
+119
-6
@@ -53,6 +53,44 @@ export interface Step {
|
||||
* 两者要同时展示(问什么 + 查到什么),合成一个字段就只能二选一 */
|
||||
result?: string
|
||||
children?: Step[]
|
||||
/** 协作团队的泳道数据。只有 supervisor 节点的顶层 Step 会有 */
|
||||
team?: TeamRun
|
||||
}
|
||||
|
||||
/**
|
||||
* 协作团队一轮里的一个人。
|
||||
*
|
||||
* ms 是**各自量出来**的耗时,不是拿整轮墙钟摊的——界面要显示"并行省了多少",
|
||||
* 那个数按 N×墙钟 推算会把快的那个也记成最慢那条,省下的时间就被夸大了。
|
||||
*/
|
||||
export interface TeamMember {
|
||||
agent: string
|
||||
instruction: string
|
||||
ms: number
|
||||
status: StepStatus
|
||||
result?: string
|
||||
}
|
||||
|
||||
export interface TeamRound {
|
||||
round: number
|
||||
/** 这一轮同时派了几个人。1 就是串行的一步 */
|
||||
parallel: number
|
||||
/** 这一轮实际花的时间:并发时是最慢那个 */
|
||||
wallMs: number
|
||||
/** 各人耗时之和:并发时它大于 wallMs,差值就是省下的 */
|
||||
sumMs: number
|
||||
reason?: string
|
||||
members: TeamMember[]
|
||||
}
|
||||
|
||||
export interface TeamRun {
|
||||
/** 花名册,按第一次出现的顺序——泳道的行顺序 */
|
||||
members: string[]
|
||||
rounds: TeamRound[]
|
||||
/** 并行一共省下多少毫秒。全程串行则为 0 */
|
||||
savedMs: number
|
||||
/** 协作是不是已经收尾了 */
|
||||
finished: boolean
|
||||
}
|
||||
|
||||
// 这两个是流式增量,后端根本不落库(_EPHEMERAL)。单次运行的 delta 量级会
|
||||
@@ -297,6 +335,28 @@ export function decodeRun(events: RunEvent[]): Step[] {
|
||||
const pendingLlm = new Map<string, Step>()
|
||||
/** supervisor 里待配对的单个 agent 步骤 */
|
||||
const pendingAgents = new Map<string, Step>()
|
||||
/** node_id → 协作团队的泳道数据。边解码边攒,最后挂到该节点的顶层 Step 上 */
|
||||
const teams = new Map<string, TeamRun>()
|
||||
|
||||
const teamOf = (nodeId: string | undefined): TeamRun | null => {
|
||||
if (!nodeId) return null
|
||||
let t = teams.get(nodeId)
|
||||
if (!t) {
|
||||
t = { members: [], rounds: [], savedMs: 0, finished: false }
|
||||
teams.set(nodeId, t)
|
||||
}
|
||||
return t
|
||||
}
|
||||
|
||||
const roundOf = (team: TeamRun, round: number): TeamRound => {
|
||||
let r = team.rounds.find((x) => x.round === round)
|
||||
if (!r) {
|
||||
r = { round, parallel: 1, wallMs: 0, sumMs: 0, members: [] }
|
||||
team.rounds.push(r)
|
||||
team.rounds.sort((a, b) => a.round - b.round)
|
||||
}
|
||||
return r
|
||||
}
|
||||
/**
|
||||
* node_id → 这个节点当前那条尚未闭合的审批步骤。
|
||||
*
|
||||
@@ -626,6 +686,18 @@ export function decodeRun(events: RunEvent[]): Step[] {
|
||||
}
|
||||
pendingAgents.set(agentKey(nodeId, d), step)
|
||||
push(step, nodeId)
|
||||
|
||||
const team = teamOf(nodeId)
|
||||
if (team) {
|
||||
const name = String(d.agent ?? '')
|
||||
if (name && !team.members.includes(name)) team.members.push(name)
|
||||
const r = roundOf(team, num(d.round) ?? 0)
|
||||
r.parallel = Math.max(r.parallel, num(d.parallel) ?? 1)
|
||||
r.members.push({
|
||||
agent: name, instruction: String(d.instruction ?? ''),
|
||||
ms: 0, status: 'running',
|
||||
})
|
||||
}
|
||||
break
|
||||
}
|
||||
|
||||
@@ -645,6 +717,23 @@ export function decodeRun(events: RunEvent[]): Step[] {
|
||||
detail: preview,
|
||||
}, nodeId)
|
||||
}
|
||||
|
||||
const team = teamOf(nodeId)
|
||||
if (team) {
|
||||
const r = roundOf(team, num(d.round) ?? 0)
|
||||
const name = String(d.agent ?? '')
|
||||
const ms = num(d.duration_ms) ?? 0
|
||||
const m = r.members.find((x) => x.agent === name && x.status === 'running')
|
||||
?? (r.members.push({ agent: name, instruction: '', ms: 0, status: 'running' }),
|
||||
r.members[r.members.length - 1])
|
||||
m.ms = ms
|
||||
m.status = 'done'
|
||||
m.result = preview || undefined
|
||||
// 这一轮实际花的是最慢那个,各人之和减去它就是省下的
|
||||
r.wallMs = Math.max(...r.members.map((x) => x.ms))
|
||||
r.sumMs = r.members.reduce((acc, x) => acc + x.ms, 0)
|
||||
team.savedMs = team.rounds.reduce((acc, x) => acc + Math.max(0, x.sumMs - x.wallMs), 0)
|
||||
}
|
||||
break
|
||||
}
|
||||
|
||||
@@ -656,17 +745,33 @@ export function decodeRun(events: RunEvent[]): Step[] {
|
||||
// 就是它,和普通 info 日志区分得开
|
||||
if (level === 'info' && d.round != null) {
|
||||
const text = String(d.message ?? '')
|
||||
const round = (num(d.round) ?? 0) + 1
|
||||
// 优先读结构化字段。以前只有一句中文消息串,这里得拿正则去拆
|
||||
// 「调度 → X(理由)」——后端文案一改就散架,而文案是会改的。
|
||||
// 老运行的事件没有这些字段,正则那条路留着兜底
|
||||
const structured = Array.isArray(d.agents)
|
||||
const m = text.match(/^调度\s*→\s*([^((]+)[((](.*)[))]\s*$/)
|
||||
const legacyTarget = m?.[1].trim() ?? ''
|
||||
const agents: string[] = structured
|
||||
? (d.agents as unknown[]).map(String)
|
||||
: (legacyTarget && legacyTarget !== 'FINISH' ? [legacyTarget] : [])
|
||||
const done = structured ? !!d.done : legacyTarget === 'FINISH'
|
||||
const reason = structured ? String(d.reason ?? '') : (m?.[2].trim() ?? '')
|
||||
const team = teamOf(nodeId)
|
||||
if (team) {
|
||||
if (done) team.finished = true
|
||||
else if (reason) roundOf(team, num(d.round) ?? 0).reason = reason
|
||||
}
|
||||
|
||||
// FINISH 是协议里的收尾标记,不是某个 agent。"交给 FINISH"
|
||||
// 会让人以为还有个叫 FINISH 的成员
|
||||
const target = m?.[1].trim() ?? ''
|
||||
const round = Number(d.round) + 1
|
||||
const title = done ? `第 ${round} 轮:结束协作`
|
||||
: agents.length > 1 ? `第 ${round} 轮:${agents.join('、')} 同时进行`
|
||||
: agents.length === 1 ? `第 ${round} 轮:交给 ${agents[0]}`
|
||||
: text
|
||||
push({
|
||||
id: `sv-${seq}`, seq, kind: 'branch', nodeId, status: 'done',
|
||||
title: !m ? text
|
||||
: target === 'FINISH' ? `第 ${round} 轮:结束协作`
|
||||
: `第 ${round} 轮:交给 ${target}`,
|
||||
detail: m ? m[2].trim() : undefined,
|
||||
title, detail: reason || undefined,
|
||||
}, nodeId)
|
||||
break
|
||||
}
|
||||
@@ -778,6 +883,14 @@ export function decodeRun(events: RunEvent[]): Step[] {
|
||||
}
|
||||
|
||||
// 还没收到 end 的工具/模型调用保持 running——它们正在进行,不是丢了
|
||||
// 泳道挂到 supervisor 节点的顶层 Step 上。节点内部那些"第N轮交给谁""X:指令"
|
||||
// 仍然留着——泳道给的是一眼看清的形状,那些步骤给的是内容,两者不重复
|
||||
for (const [nodeId, team] of teams) {
|
||||
if (!team.rounds.length) continue
|
||||
const step = nodeSteps.get(nodeId)
|
||||
if (step) step.team = team
|
||||
}
|
||||
|
||||
return out
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user