Repository navigation
register-notifies-watchers 的「订阅者重读看得见新 body」用例，实际钉住的是 get() 里的 microtask 深度，不是它声称的顺序保证 #6043
Description
做 #5840 时被这条用例真实拦下一次，按 PD #10 单独记录。观察类：今天没有用户碰得到的行为缺陷，红的是测试的健壮性，不是产品。
事实
packages/metadata/src/register-notifies-watchers.test.ts：
it('announces AFTER the write lands, so a subscriber that re-reads sees the new body', async () => {
let readBack: unknown;
manager.subscribe('object', async () => {
readBack = await manager.get('object', 'account');
});
await manager.register('object', 'account', { name: 'account', label: 'Fresh' });
expect(readBack).toEqual({ name: 'account', label: 'Fresh' });
});notifyWatchersLocal 用 void callback(event) 派发（metadata-manager.ts，private notifyWatchersLocal 一带），从不 await 回调。所以 register() 返回时，订阅者那句 await manager.get(...) 还没结算；readBack 能在断言前被赋值，靠的是「get() 内部的 await 跳数」恰好少于 await manager.register(...) 让出的微任务数。
为什么这是个 finding
它声称钉的顺序保证（先写落地、后广播）是真的，也确实值得钉。但这个写法测不到它：真正决定成败的是 get() 里有几个 await 帧。
实测证据（#5840 开发过程中）：把 get() 从
async get(type, name) { /* registry 命中直接 return */ ... const result = await this.load(type, name); ... }改写成等价的 return (await this.getDiagnosed(type, name)).data; —— 语义逐字不变、registry 优先不变、返回值不变，只多了一个 async 帧 —— 这条用例即刻变红，readBack 为 undefined。#5840 因此放弃了那个（更整洁的）委托写法，保留三行重复以维持逐帧一致，并把原因写进了 get() 的 TSDoc。
也就是说：一次纯粹的等价重构会让它红，而它声称保护的顺序保证毫发无损。反过来，若真把广播挪到写入之前（它要防的那个回归），这条用例大概率照样绿——订阅者读到的仍是同一个 registry map，只是内容早晚问题，readBack 该有值还是有值。假阴性与假阳性都在。
可能的修法（未择，留给承接座位）
- (a) 让订阅者把 promise 记下来，断言前
await它——最小改动，测的仍是「重读看得见」，但不再依赖跳数； - (b) 直接钉真正的不变量：在回调里同步读
(manager as any).registry.get('object').get('account')，断言广播那一刻写已落地——这才是「AFTER the write lands」的字面意思； - (c) 顺带考虑
notifyWatchers是否该给一个「已排空」的可观测点，供测试与集群桥接共用（范围更大，多半不值）。
倾向 (b)：它把断言对准了方法名承诺的那件事，且对 get() 的实现细节完全免疫。
Activity
claude commentedon Aug 6, 2026
发现分诊(#4949 纪律):晋级 —— 摘
finding,改pm:queue+domain:engine-core过时前提检查(
origin/main9e3709a):packages/metadata/src/register-notifies-watchers.test.ts仍在;metadata-manager.ts的notifyWatchersLocal(:2665)确实以void callback(event)(:2671)派发、从不 await,两个调用点在:2643/:2748⇒ 立单事实全部成立。域:落点包
packages/metadata⇒domain:engine-core(按域表,不按「测试」词汇归 devx)。为什么晋级而不是持有:三条都满足自动入队的形状 —— 落点唯一、修法已点名(倾向 (b):回调里同步读 registry,断言「广播那一刻写已落地」)、且是 test-only 改动。更关键的是它已经产生了真实成本:#5840 因这条用例放弃了等价且更整洁的
getDiagnosed委托写法,保留三行重复,即一条测试的偶然实现细节正在反向约束产品代码的形状。假阴性(把广播挪到写入前它照样绿)与假阳性(纯等价重构即红)同时存在,属「闸门没有钉住它声称钉住的东西」一族,与本轮前例 #5916 同形。串行(硬):在飞 PR #6051(#5840)改的是同包
metadata-manager.ts(实测 diff 命中,但不含本单的测试文件)⇒ 无文件冲突,但顺序有意义:本单应排在 #6051 落地之后,并在同一次改动里复核该 PR 留在get()TSDoc 里的常驻说明(「为保持逐帧一致故不委托」)是否随本单修好而过期 —— 若过期而不改,下一个作者会照着一条已失效的理由继续保留重复代码。范围:⛔ 只做 (b)(必要时加 (a));正文的 (c)(给
notifyWatchers增设「已排空」可观测点)明确划在范围外,它是接口扩展,需另行取 appetite。本评论来自分诊座位 Routine(#5474 试点),不构成认领。
Generated by Claude Code
hotlong commentedon Aug 7, 2026
迁移:
domain:engine-core→domain:metadata(维护者 2026-08-07 批准拆分;座位贴 #6367,SKILL PR #6370)。落点证据:packages/metadata的 register-notifies-watchers.test.ts 用例实际钉住MetadataManager.get的 microtask 深度而非声称的顺序保证 —— metadata 通知接缝。全量分类审计 49 卡三判,本卡 MOVE/high。分诊会话:session_01BickTBKm2JYSNnrtPT8ysa。Generated by Claude Code
baozhoutao commentedon Aug 7, 2026
Claim: PM loop round 1 (
domain:metadataseat, sticker #6367)Session:
session_01KDU3qAuJyajAQm3GkUXfdABranch:
claude/issue-6043-watcher-test-real-invariantWorktree:
objectstack-issue-6043Domain:
domain:metadataFile surface:
packages/metadata/src/register-notifies-watchers.test.ts+ theget()TSDoc block inpackages/metadata/src/metadata-manager.ts(comment-only touch, per the triage promotion comment's instruction to re-verify the note PR #6051 left there). Stop on breach; explain in the report.Serial constraints cleared: PR #6051 (#5840) — the hard predecessor named at triage — merged 2026-08-07 01:49Z; no open claim touches these files. #5309 (future round, also lands on metadata-manager.ts) deliberately sequenced after this card.
Container class: S —
mode:subagent, shared container.Generated by Claude Code
baozhoutao commentedon Aug 8, 2026
Correction to the claim above: the session ID contains a transposition typo. The correct ID is
session_01KDU3qAuJyajAQm3GkUXdfA(matches the seat takeover audit on #6367). Branch, worktree, and all other claim fields are unchanged.Generated by Claude Code
baozhoutao commentedon Aug 8, 2026
ACCEPT (PM review,
domain:metadataseat, sessionsession_01KDU3qAuJyajAQm3GkUXdfA) — PR #6547.Shipped: the "announces AFTER the write lands" case now pins the ordering guarantee itself — ruling option (b) (synchronous registry peek inside the watcher callback, with an overwrite variant so an early broadcast fails readably as
V1-vs-V2) plus (a) (the subscriber'sget()promise captured and awaited). PR #6051's standing TSDoc note onget()is rewritten to current reality (comment-only;get()not refactored — correctly left as a follow-up suggestion).Review verified against GitHub: two files exactly on the claimed surface; the
metadata-manager.tshunk is comment-only (read line-by-line); ESLint and TypeScript Type Check jobs bothsuccess, 30 checks zero failures. Reverse verification measured both failure modes with directions fixed before running, and honestly falsified the issue's own prediction — the old case does go red on an early broadcast, but with anundefinedindistinguishable from a harmless refactor; the one-notch-finer true false negative (hoisting past the realtime publish + writable-loader save loop stays 15/15 green) is recorded and filed as #6548 (finding). Tests + comment only ⇒skip-changeset(on and read back).PR flipped ready and queued (auto-merge); the issue closes on merge via the
Fixesline.Generated by Claude Code