fix(subagent): 子代理页实时预览改为批量接口,避免 O(N²) history 请求风暴 (#298)

* fix(subagent): 子代理页实时预览改为批量接口,避免 O(N²) history 请求风暴

## 背景

DSH Web UI 中,当主代理同时运行多个子代理时,`dsh-better-sidebar` 的「任务管理/子代理」页面会出现明显卡顿。经调查,卡顿根因不在页面渲染,而在实时预览的请求模式。

## 根因

旧实现中,每个 running 子代理卡片都独立轮询 `subagents.history`:

- N 个 running 子代理 → 每 3 秒产生 N 次 `subagents.history` 请求;
- 每次 `subagents.history` 都会在 host 侧触发一次完整的 `listChildren` 全量子代理目录枚举;
- 子代理数量越多,服务端重复扫描越严重,近似 O(N²) 放大;
- 请求超过 3 秒时,客户端还会 abort 旧请求并重发,进一步形成请求风暴。

## 改动内容

1. **新增 host 批量接口 `POST /sidebar/api/subagents.live`**
   - 请求 `{ rootSessionId }`;
   - 通过 `ctx.subagents.listDescendants(rootSessionId)` 一次枚举整棵子代理树;
   - 仅对 running 且非 Side Chat 的子代理,从 session events 尾部提取最新 `text/tool` 活动;
   - 返回 `{ live: Record<childId, { text?, tool? }> }`;
   - 缺 `subagents`/`agents` 服务或枚举失败时返回 503;单个子代理日志异常只跳过该子代理。

2. **客户端改为单轮询 + 单飞行请求**
   - 删除每个 running 子代理独立的 `subagents.history` 定时器;
   - 在 `SubagentView` 中合并为一个 `useSubagentLive` 轮询 hook;
   - 使用递归 `setTimeout` 和单在途请求,慢请求不再被 abort/重发;
   - 页面隐藏或 root 变化时停止轮询并中止在途请求。

3. **解析器优化**
   - `lastActivity` 改为接收原始 `SidebarSessionEvent[]`;
   - 从事件尾部向前扫描,找到最新 text 和 tool 后提前停止,避免复制/遍历完整历史。

4. **UI 保持不变**
   - 实时预览的展示逻辑、文案、截断规则均未改变。

## 测试

- 新增 `tests/subagent-live-route.spec.ts`:覆盖批量接口成功聚合、只返回 running、跳过 Side Chat、缺服务 503、枚举失败 503、`ctx.get` 回退、bad-request。
- 新增 `tests/subagent-live-polling.spec.tsx`:覆盖每轮只发 1 次 `subagents.live`、不再调用 `subagents.history`、单在途不并发。
- 更新 `tests/subagent-activity.spec.ts` 与 `tests/subagent-jobs-view.spec.tsx`。
- `pnpm typecheck` ✅
- `pnpm build` ✅
- 相关测试 29 个全部通过 ✅

> 本机完整 `pnpm test` 仍有与本改动无关的环境性失败(node-pty 超时、Windows CRLF 断言、pty-deps 路径断言),不影响本 PR 逻辑。

* fix(subagent): live 预览以 catalog activity 为准,删除 agents 第二事实源 (#298)

- host 路由 running 门改回 entry.activity === 'running'(与客户端卡片渲染门一致),
  恢复 PR 承诺的 UI 不变:turn 间 idle 的 continuable 子代理仍显示最后活动
- 删除 ctx.agents 依赖与 SidebarAgent.status 新增(agents 缺失时不再静默 200 空 map;
  503 契约收窄为只涉及 subagents,与文档/测试一致)
- service 解析统一为 ctx.get('subagents')(兄弟路由惯例;插件未 inject subagents,
  直接属性访问有抛错风险)
- buildSubagentLiveApi 签名改收完整 Context,删除 SidebarSubagentLiveContext 窄接口
- api.ts 复用 LastActivity,删除重复类型 SubagentLiveEntry
- 测试:route spec 删除 agents 相关用例;polling spec 新增页面隐藏停止轮询、
  root 变化清空 live 并重启轮询两个用例

* fix(subagent): live 预览恢复最近 12 条消息窗口,排除陈旧活动 (#298)

CR 反馈:新批量接口把完整 session event log 喂给 lastActivity,
- 长时间只输出文本、早期调用过工具的 running 子代理会持续显示陈旧 tool
- 无 tool 的长会话每次轮询反向扫描整个事件数组,重新引入与历史长度
  相关的 CPU 成本

修复(与旧 subagents.history({ maxMessages: 12 }) 窗口语义一致):
- lastActivity 增加可选 maxMessages 参数:反向扫描按表面消息
  (user/message + assistant/message)计数,窗口满后更老事件一律跳过
- 路由传 LIVE_WINDOW_MESSAGES = 12(导出常量)
- 回归测试:窗口外旧 tool 不返回;窗口内 tool 保留;非消息事件不消耗配额

---------

Co-authored-by: Menghuan1918 <menghuan2003@outlook.com>
This commit is contained in:
NuperAki
2026-08-21 18:15:49 +08:00
committed by GitHub
co-authored by Menghuan1918
parent a9d3451415
commit 6cea4bdd6f
12 changed files with 815 additions and 143 deletions
+70 -66
View File
@@ -41,7 +41,7 @@ import {
isSideThreadSummary,
rootAncestor,
} from './subagent-detect.ts'
import { lastActivity } from './subagent-activity.ts'
import { type LastActivity } from '../subagent-activity.ts'
import { SIDE_LABEL_PREFIX } from '../sidechat-core.ts'
import {
collectTreeJobs,
@@ -150,56 +150,14 @@ function CatalogLoadingRows(props: {
}
/**
* The live lines of one RUNNING subagent card: the last text output and the
* last tool call of the child's history tail, refreshed every few seconds
* while the page is visible. Idle cards render nothing (a quiet topology); a
* running child with neither output yet reads "thinking…".
* The live lines of one RUNNING subagent card: a pure presentation of the
* batch `subagents.live` activity. The polling lives in one place (the
* SubagentView hook), not per card. A running child with neither output yet
* reads "thinking…".
*/
function SubagentLiveLines(props: {
ctx: Context
parentSessionId: string
childSessionId: string
mode: SidebarSubagentAddress['mode']
running: boolean
/** The page is visible (active tab + open panel): skip polling otherwise. */
active: boolean
}) {
const { ctx, parentSessionId, childSessionId, mode, running, active } = props
const [live, setLive] = useState<ReturnType<typeof lastActivity>>({})
const controllerRef = useRef<AbortController | undefined>(undefined)
const address = useMemo(
() => ({ parentSessionId, childSessionId, mode }),
[parentSessionId, childSessionId, mode],
)
const load = useCallback(async (): Promise<void> => {
controllerRef.current?.abort()
const controller = new AbortController()
controllerRef.current = controller
try {
const response = await ctx.connection.api.subagents.history(
{ ...address, maxMessages: 12 },
controller.signal,
)
if (!response.result.ok) return
setLive(lastActivity(response.result.value.events))
} catch {
// Aborted by a newer pull or a wire failure: keep the last known lines.
}
}, [ctx, address])
useEffect(() => {
if (!active) return
void load()
if (!running) return
const timer = window.setInterval(() => { void load() }, POLL_MS)
return () => { window.clearInterval(timer) }
}, [load, running, active])
useEffect(() => () => { controllerRef.current?.abort() }, [])
if (!running) return null
if (live.text === undefined && live.tool === undefined) {
function SubagentLiveLines(props: { live: LastActivity | undefined }) {
const { live } = props
if (live?.text === undefined && live?.tool === undefined) {
return <span className={css.subagentLive}>{t('subagentThinking')}</span>
}
return (
@@ -219,6 +177,59 @@ function SubagentLiveLines(props: {
)
}
/**
* One shared live-preview poller for the whole Subagent tree. Unlike the old
* per-card `subagents.history` timers, this sends at most ONE `subagents.live`
* request at a time: a recursive timeout starts only after the previous
* request settles, so a slow host never sees abort/restart storms.
*/
function useSubagentLive(
rootId: string | undefined,
active: boolean,
): Readonly<Record<string, LastActivity>> {
const [live, setLive] = useState<Record<string, LastActivity>>({})
const controllerRef = useRef<AbortController | undefined>(undefined)
// A new tree must never inherit another root's live previews.
useEffect(() => { setLive({}) }, [rootId])
useEffect(() => {
if (rootId === undefined || !active) return
const targetRootId = rootId
let disposed = false
let timer: number | undefined
const schedule = (): void => {
if (disposed) return
timer = window.setTimeout(() => { void load() }, POLL_MS)
}
async function load(): Promise<void> {
if (disposed) return
const controller = new AbortController()
controllerRef.current = controller
try {
const result = await api.subagentsLive(targetRootId, controller.signal)
if (!disposed) setLive(result.live)
} catch {
// Keep the last known live map; the next scheduled poll retries.
} finally {
if (controllerRef.current === controller) controllerRef.current = undefined
if (!disposed) schedule()
}
}
void load()
return () => {
disposed = true
if (timer !== undefined) window.clearTimeout(timer)
controllerRef.current?.abort()
controllerRef.current = undefined
}
}, [rootId, active])
return live
}
interface RowsProps {
parentSessionId: string
catalog: SidebarSubagentCatalog | undefined
@@ -227,16 +238,15 @@ interface RowsProps {
level: number
/** The currently-open session id (highlighted in the topology). */
currentSessionId: string
/** The page is visible (active tab + open panel): live polling pauses otherwise. */
active: boolean
ctx: Context
/** The batch live-preview map (child id → latest activity). */
live: Readonly<Record<string, LastActivity>>
openChild: (address: SidebarSubagentAddress) => void
refresh: (parentSessionId: string) => void
}
/** Render one topology level; branches are always expanded (lazy catalogs). */
function CatalogRows({
parentSessionId, catalog, catalogs, byId, level, currentSessionId, active, ctx,
parentSessionId, catalog, catalogs, byId, level, currentSessionId, live,
openChild, refresh,
}: RowsProps) {
const emptyLoading = catalog?.state === 'loading' && catalog.entries.length === 0
@@ -327,14 +337,9 @@ function CatalogRows({
<span className={css.subagentContent}>
<span className={css.subagentLabel}>{label}</span>
<span className={css.subagentSecondary}>{secondary}</span>
<SubagentLiveLines
ctx={ctx}
parentSessionId={parentSessionId}
childSessionId={entry.id}
mode={entry.mode}
running={entry.activity === 'running'}
active={active}
/>
{entry.activity === 'running' && (
<SubagentLiveLines live={live[entry.id]} />
)}
</span>
</div>
{!knownLeaf && (
@@ -355,8 +360,7 @@ function CatalogRows({
byId={byId}
level={level + 1}
currentSessionId={currentSessionId}
active={active}
ctx={ctx}
live={live}
openChild={openChild}
refresh={refresh}
/>
@@ -658,6 +662,7 @@ export function SubagentView(props: {
const rootId = useMemo(() => rootAncestor(byId, sessionId), [byId, sessionId])
const rootCatalog = rootId === undefined ? undefined : catalogs[rootId]
const rootSummary = rootId === undefined ? undefined : byId[rootId]
const live = useSubagentLive(rootId, active)
/** Catalog owners currently consuming live membership updates. */
const observedRef = useRef(new Set<string>())
@@ -852,8 +857,7 @@ export function SubagentView(props: {
byId={byId}
level={1}
currentSessionId={sessionId}
active={active}
ctx={ctx}
live={live}
openChild={openChild}
refresh={refresh}
/>
+11
View File
@@ -7,6 +7,7 @@
* request). Failures surface as {@link SidebarApiError} with the wire code.
*/
import { encodeHtmlUrl } from '../html-route.ts'
import type { LastActivity } from '../subagent-activity.ts'
import type { SidechatThreadInfo } from '../sidechat-core.ts'
import type { BrowserProbeResult } from './browser.ts'
@@ -80,6 +81,9 @@ export interface JobOutputResult {
read: boolean
}
/** The `subagents.live` response: running child id → latest activity. */
export type SubagentLiveResult = { live: Record<string, LastActivity> }
/** Terminal dependency status (mirror of the host's depsStatus; issue #140). */
export type TerminalDepsStatus =
| { ok: true }
@@ -245,6 +249,13 @@ export const api = {
id,
...(reason !== undefined ? { reason } : {}),
})),
/**
* One batch live-preview fetch for the whole Subagent tree. The payload is
* the already-resolved topology ROOT (not a session scope); the host
* enumerates descendants once and folds running children's activity.
*/
subagentsLive: (rootSessionId: string, signal?: AbortSignal) =>
call<SubagentLiveResult>('subagents.live', { rootSessionId }, signal),
/** Create a Side Chat thread: a child session seeded with the parent's
* full log up to now. Empty question = immediate create (Codex-style):
* the thread opens empty, the first prompt carries the boundary. */
-67
View File
@@ -1,67 +0,0 @@
/**
* Pure derivation of the compact LIVE line shown on a running subagent card:
* the last text output and the last tool call of the child's history tail
* (`subagent.history` events, oldest → newest). Renders nothing itself — the
* SubagentLiveLine component turns this into the card's status lines.
* Kept framework-free so the parser is unit-testable in the node environment.
*/
import type { SidebarHistoryEntry } from '../context-types.ts'
/**
* Extract the concatenated plain text of a content-block list (the durable
* `ContentBlock[]` shape, structurally: blocks with `type: 'text'` carry
* `text`; anything else — tool_use, image, … — contributes nothing).
* @param content - the raw `content` field of a message event.
* @returns the joined text, or undefined when the message carries no text.
*/
export function contentText(content: unknown): string | undefined {
if (!Array.isArray(content)) return undefined
const parts: string[] = []
for (const block of content) {
if (block === null || typeof block !== 'object') continue
const candidate = block as { type?: unknown; text?: unknown }
if (candidate.type === 'text' && typeof candidate.text === 'string') {
parts.push(candidate.text)
}
}
return parts.length > 0 ? parts.join('\n') : undefined
}
/** The live status of one subagent card (both fields optional). */
export interface LastActivity {
/** The latest assembled assistant text output in the tail. */
text?: string
/** The latest tool call in the tail. */
tool?: { name: string; args: string }
}
/**
* Fold a history tail into the last text output + last tool call (each is
* the LAST occurrence in event order). Lifecycle events and raw
* `assistant/chunk` rows are ignored — the card shows what the subagent is
* doing right now, not its plumbing.
* @param entries - the tail page from `subagent.history` (oldest → newest).
* @returns the last text and/or tool call; an empty object when the tail has neither.
*/
export function lastActivity(entries: SidebarHistoryEntry[]): LastActivity {
let text: string | undefined
let tool: { name: string; args: string } | undefined
for (const entry of entries) {
const { type, data } = entry.event
if (type === 'assistant/message') {
const message = data.message as { content?: unknown } | undefined
const extracted = contentText(message?.content)
if (extracted !== undefined) text = extracted
} else if (type === 'tool/call') {
tool = {
name: typeof data.name === 'string' ? data.name : 'tool',
args: typeof data.arguments === 'string' ? data.arguments : '',
}
}
}
if (text === undefined && tool === undefined) return {}
return {
...(text === undefined ? {} : { text }),
...(tool === undefined ? {} : { tool }),
}
}