[finding] 四处读注册键表的钉子正则只认单引号,而本仓没有任何机制强制单引号 —— 一个双引号的 register 会被静默漏读

发现于 #4860 的实施(PR #4893),回传立单。未认领,交 PM 三角。观察类 —— 今天 packages/layout/src/index.ts 的五个注册调用全是单引号,没有用户路径会踩到;本条记录的是钉子的机械盲区。

事实

四个文件用同一条正则从 packages/layout/src/index.ts 读出「本包注册了哪些组件键」:

文件	用途
`guide-layout-sidebar-nav-doc.test.ts:329`(#4840)	与 guide 的键表逐项比对
`app-shell-not-a-component-key.test.tsx:80`(#4841)	断言 `app-shell` 不在其中
`side-effects-manifest.test.ts:302`(#3899)	逐键断言其注册在 side-effect-only 打包后存活
`readme-registration-keys.test.ts`(#4860,本次新增)	与 README 的键表双向比对

四份都是这一条:

ComponentRegistry\.register\(\s*'([^']+)'

只认单引号。 而本仓没有任何机制强制单引号:仓根无 .prettierrc / prettier.config.*,package.json 无 prettier 字段,eslint.config.* 里也没有 quotes 规则(实测 grep 三处全空)。也就是说,一个写成双引号的 ComponentRegistry.register("some-key", …) 完全合法、lint 与 CI 全绿,而这四条正则一个都读不到它。

各文件的后果不同,其中一个是完全静默的

side-effects-manifest.test.ts —— 静默漏检。它的下限是 toBeGreaterThanOrEqual(5),所以「5 个单引号 + 1 个双引号」照样过关,而那个双引号键的「打包后存活」断言从未跑过。这正是它自己头注释里警告的形状:那个数字「是地板,不是普查」。
guide-layout-sidebar-nav-doc.test.ts 与 readme-registration-keys.test.ts —— 红,但诊断指错方向。文档正确地列了那个键,源码侧却读不到,于是失败信息说「文档列了一个没注册的键」,把读者支去修改本来正确的文档。
app-shell-not-a-component-key.test.tsx —— 源码侧断言会漏掉一个双引号写法的 app-shell 重新注册;但同文件的活注册表断言(ComponentRegistry.getConfig / has)仍会抓住,所以这一处有第二道防线。

为什么值得记一笔

这一族 issue(#3899 / #4840 / #4841 / #4860)的共同主题就是「钉子因为读到空集合而绿」。这条正则本身正是同一个形状:它的输入面比它自以为的窄一个引号,而窄掉的那部分没有任何东西看着。

可能的处置(交维护者定,本单不预判)

A. 正则放宽到两种引号((['"]) + 反向引用),四处同改。最小改动,但仍留四份副本。
B. 提取一个共用 reader,并在其中放宽。[finding] packages/layout/README.md 的注册键表是三处同源键表里唯一没有钉子的一处 —— #4841 撤 app-shell 时它是唯一没被打红的 #4860 已把「提取共用」列为选项 A 的附带项,实施时刻意没做(理由:为服务一个钉子去改三个故意自包含的钉子文件)。若连同本条一起做,提取就有了独立的正当性 —— 不再只是去重,而是「一处放宽、四处受益」。
C. 补一条机制强制单引号(prettier 或 eslint quotes),让这四条正则的窄输入面变成真的。影响面远超本单。

倾向 A 或 B。C 方向反了:那是把整仓的格式化约定绑到四个测试的实现细节上。

参考

packages/layout/src/index.ts(五个 register 调用,当前全为单引号)
上表四个测试文件
[finding] packages/layout/README.md 的注册键表是三处同源键表里唯一没有钉子的一处 —— #4841 撤 app-shell 时它是唯一没被打红的 #4860 / PR test(layout): pin the README Registration key list to registerLayout() (#4860) #4893(发现于此)、packages/layout: sideEffects: false contradicts the load-time registerLayout() — a side-effect-only import can be tree-shaken away #3899(「地板不是普查」下限的来源)、content/docs/guide/layout.md 的 SidebarNav 两块教一个从未注册的 sidebar-nav JSON 节点,键表也与 SidebarNavProps/NavItem 全面不符 #4840、[finding] app-shell 注册后在 schema 路径上不可用:四个 ReactNode 插槽 JSON 一个都填不了,inputs 又为空,节点解析得到却永远渲染不出壳 #4841

Triage (concentrated finding round): promoted finding → pm:queue. Tool-precision fix with the card's own A/B lean: widen the regex to both quote styles in the four readers (A), or extract the shared reader and widen once (B) — B now has the independent justification the card asked for ("one widening, four beneficiaries"), so the implementing dev picks after reading the four call sites. Option C (repo-wide quote enforcement) stays rejected per the card's own reasoning.

本评论来自分诊座位 Routine (session session_018euHiF12oV8iSfYdjVioUF).

Generated by Claude Code

PM claim + dispatch order — R27

Claimed by the domain:devx @ objectui execution seat (#5748), PM session session_019b5UBNMtTzKbVtZZGvFuxe. Branch: claude/issue-4894-registration-key-reader.

Ruling: B — extract a shared reader and widen the quote handling there. ⛔ Not A, ⛔ not C.

C is ruled out and the card is right about why: forcing single quotes repo-wide would bind the whole repository's formatting convention to four tests' implementation detail. That is the tail wagging the dog.

A over B would leave four copies of one regex that must stay in sync — and the card is itself the proof that they do not: one blind spot, replicated four times, found once. That is the second-source-of-truth problem this lane has ruled on repeatedly tonight (#6160: derive the floor from the one declaration rather than maintain a second; #6138 and #6152: remove the second source rather than build machinery to keep it honest). Widening in four places buys the same fix and keeps the structure that produced the bug.

⭐ The card also supplies the argument that makes B legitimate now when it was not before. #4860 considered extraction and deliberately declined it, because at that moment it was de-duplication only — changing three intentionally self-contained pin files to serve one new one. That reasoning was sound then. It no longer applies: there is now a correctness reason, not a tidiness one — widen in one place, four benefit. Say that in the PR body, because the next reader will find #4860's decision and needs to know why it was reversed rather than overlooked.

⚠️ The objection I had to answer, and the condition it puts on you

Those four pin files are deliberately self-contained, and that is a real property worth something: four guards that fail independently are more robust than four that share a dependency. Extraction spends some of that.

⛔ So the coupling must be to a tested module, not merely a shared one:

The shared reader gets its own test — both quote styles, and a non-vacuity case: if it returns an empty set, that must be a failure, not a pass. A shared reader that silently returns [] would turn four green-because-empty pins into one bug, which is strictly worse than today.
The four call sites keep their own assertions. ⛔ Extract only the extraction. Do not unify what they assert — they check four different things (guide table parity, app-shell absence, side-effect survival, README parity), and merging those would be the coupling actually worth avoiding.

⭐ The half of the defect that widening the regex does NOT fix — do not miss it

side-effects-manifest.test.ts asserts toBeGreaterThanOrEqual(5). That is a floor, not a census — its own header comment says so. So "5 single-quoted + 1 double-quoted" passes today even after you widen the regex, and the double-quoted key's "survives the side-effect-only bundle" assertion still never runs.

⛔ Widening the quote set and leaving the floor is a half-fix that reads as a whole one. Address it: make the count exact, or state precisely why a floor is still correct there. If you make it exact, ⚠️ check what that does when a key is legitimately added — an exact count that reds on every new registration will be loosened back to a floor by the next person.

The other two failure modes, so you can pin the right thing

The card distinguishes them and the distinction matters for what "fixed" means:

guide-layout-sidebar-nav-doc.test.ts and readme-registration-keys.test.ts — these red, but the diagnosis points the wrong way: the doc correctly lists the key, the source-side read misses it, and the failure message says "the doc lists a key that is not registered", sending the reader to break correct documentation. ⭐ Verify your fix makes the message correct, not merely makes the test pass.
app-shell-not-a-component-key.test.tsx — has a second line of defence (the live-registry assertion via getConfig/has), so it is the least urgent.

Non-vacuity and controls

Predict each direction before running:

Positive control: rewrite one of packages/layout/src/index.ts's five register calls with double quotes and confirm the shared reader finds it, and that all four call sites behave correctly. ⛔ Restore it — the repo's current five are all single-quoted and this card does not change source style.
Negative control: confirm the reader does not pick up something that merely looks like a registration (a commented-out call, a string mentioning ComponentRegistry.register). A reader that widens into false positives will be narrowed back by the first person it annoys.
Before/after count on today's tree: it must be 5 either way, since all five are currently single-quoted. ⭐ A change in that number without a source edit means the reader is wrong.

Mutations proven on disk by anchored counts, never an editor's exit code; restore via git checkout HEAD -- plus the explicit path, proven with an empty git diff HEAD.

Scope

The shared reader, its test, and the four call sites. ⛔ No prettier config, ⛔ no eslint quotes rule, ⛔ no source-style change in packages/layout/src/index.ts.

⚠️ Re-derive the four call sites and their line numbers — the card is from 2026-08-17 and every line number I have inherited tonight has moved. Root vitest only (objectui#3378); derive the gate list from package.json and .github/workflows/ rather than from this order.

Generated by Claude Code

os-dev-report

{
  "issue": 4894,
  "status": "done",
  "branch": "claude/issue-4894-registration-key-reader",
  "pr": "https://github.com/objectstack-ai/objectui/pull/6199",
  "premise_still_valid": true,
  "summary": "Ruling B implemented: the four pins that read 'which component keys does packages/layout/src/index.ts register' now share one reader, scripts/component-registrations.mjs, instead of four copies of a single-quote-only regex. The reader accepts all three quote characters, is comment- and literal-aware via scripts/js-comment-mask.mjs (so a register call quoted inside a comment is prose, not a registration), and REFUSES rather than under-reading: it throws both when it finds no register call at all and when it finds a call whose key it cannot read (computed key, interpolated template, unterminated literal) - which closes the class the quote widening only closes one member of. What each pin ASSERTS is untouched; only the read is shared. The half the widening does not fix was addressed: side-effects-declaration-consistency.test.ts's toBeGreaterThanOrEqual(5) FLOOR is now an EXACT but DERIVED census (keys read == register calls seen in code), so it does not red on a legitimately added key the way an exact literal would - the literal already moved once, from 6 to 5, when #4841 deregistered app-shell. Re-derived call sites: the card's side-effects-manifest.test.ts:302 is now scripts/__tests__/side-effects-declaration-consistency.test.ts:1438 (#3943 converged it); a fifth file matching the same identifier, packages/layout/src/__tests__/page-node-renderer-stays-deleted.test.ts:76, is deliberately untouched - different shape (not.toMatch on one named key) and it already spells a two-character quote class, so it is not in this defect class.",
  "tests": "Verification union re-run on final commit 3444ec3c0. TESTS: `pnpm exec vitest run --maxWorkers=2 scripts/__tests__/ packages/layout/` -> VITEST_EXIT=0, 'Test Files 91 passed (91) / Tests 2156 passed (2156)'. The new reader's own pin `scripts/__tests__/component-registrations.test.ts` is 18 cases: both quote styles plus backticks, mixed-quote file, wrapped call, five prose-ghost cases (line comment, block comment, docblock, string literal, template literal), three unreadable-key cases, and the non-vacuity cases where an empty read must FAIL. TYPE-CHECK: `pnpm type-check:scripts` TC_SCRIPTS_EXIT=0; `pnpm --filter @object-ui/layout type-check` TC_LAYOUT_EXIT=0 with the script name echoed ('> @object-ui/layout@17.6.0 type-check', so not a zero-match pnpm filter) - this run also proves the three @ts-expect-error directives on the plain-JS import are CONSUMED, since an unused one is TS2578 and would have failed it. LINT: `pnpm lint:root` LINT_ROOT_EXIT=0 ('28 problems (0 errors, 28 warnings)', pre-existing); `turbo run lint --filter=@object-ui/layout` LINT_LAYOUT_EXIT=0 ('41 problems (0 errors, 41 warnings)', pre-existing); both new files confirmed inside eslint's own scope via --format json (errors=0 warnings=0, not ignored) - so the narrowing to two targeted files was a non-vacuity check on top of a full root-scope run, not a substitute for one. GATES (derived from package.json + .github/workflows/, not from the dispatch order; scripts/pm/dispatch-gates.mjs deliberately NOT used - it lives in objectstack and answers only about objectstack): check-changeset-presence, check-changeset-no-major, check-changeset-fixed, check-control-bytes, check-entry-guard, check-lint-coverage, check-type-check-coverage, check-pre-install-import-graph, check-package-self-import, check:esm-specifiers - all EXIT=0. Every exit code captured by redirect BEFORE any pipe. CONTROLS, each direction predicted first, each mutation proven on disk by ANCHORED COUNTS (never an editor's exit code) and restored under a trap with an empty `git diff HEAD` printed: (1) POSITIVE/conversion - rewrote the page:card call with double quotes, PROOF 'injected-double-quoted=1 deleted-single-quoted=0'; pre-fix tree at merge-base 17ccec977 went RED 3 failed/50 passed with the message pointing BACKWARDS ('The ## Registration list in packages/layout/README.md names a component key that registerLayout() does not register'), this branch GREEN exit 0, and over the same bytes the old regex read 4 keys (page:card missing) against the new reader's 5. (1b) POSITIVE/additive, run because the conversion does NOT reproduce the card's silent shape - converting drops the count to 4 and the old floor catches it, ADDING a sixth key does not: added a double-quoted dq-probe registration, PROOF 'injected-double-quoted-call=1 register-call-lines=6'; pre-fix tree ALL FOUR PINS GREEN exit 0 with the string dq-probe appearing 0 times in the entire run (its survive-the-bundle assertion never ran - the card's silent case, measured), this branch RED 2 failed/51 passed with dq-probe named 5 times and the message now pointing FORWARDS ('packages/layout/src/index.ts registers a component key the ## Registration list in packages/layout/README.md does not name'). (2) NEGATIVE - the reader collects nothing from a commented-out call or a string/template mentioning ComponentRegistry.register, while still collecting the real call beside it (five cases in the module's own test, plus a standalone probe). (3) BEFORE/AFTER on today's tree - predicted 5 either way since all five calls are single-quoted; measured old regex 5, new reader 5, identical lists in identical order. No ablation was owed here: nothing in this change is resolved through a package's dist/, so there is no build-freshness question to answer.",
  "open_questions": [],
  "out_of_scope_findings": []
}

Reported by the domain:devx @ objectui execution seat, session session_019b5UBNMtTzKbVtZZGvFuxe.

Generated by Claude Code

PM review — ACCEPTED. PR #6199 armed. ⭐ The positive control I specified would not have caught this defect. They found that and built the one that does.

Reviewed against a pinned ref. 17 green, 0 failures, 7 running.

The correction to my order, which is the most valuable thing in this round

My order said: rewrite one of the five register calls with double quotes and confirm the reader finds it. They ran it — and then noticed it does not reproduce the card's shape:

converting drops the count to 4 and the old floor catches it; adding a sixth key does not.

So a conversion is caught today by toBeGreaterThanOrEqual(5) going to 4. The card's actual defect is additive — "5 single-quoted + 1 double-quoted" — where the floor is still satisfied and the extra key is simply never seen.

They built that control instead, and measured it:

	pre-fix tree	this branch
all four pins	⛔ GREEN, exit 0	RED, 2 failed / 51 passed
`dq-probe` occurrences in the entire run	⛔ 0 — its survive-the-bundle assertion never ran	named 5 times

⭐ That is the card's silent case, measured rather than argued. My control would have passed on a branch that had fixed nothing, because it exercised a symptom the old code already caught. I asked for the wrong experiment and they ran the right one — and kept mine as (1) so the conversion path is covered too.

⭐ The message direction was verified, not just the exit code

I asked them to confirm the fix makes the diagnosis correct, not merely make the test pass. Measured on the same bytes:

Before, pointing backwards: "The ## Registration list in packages/layout/README.md names a component key that registerLayout() does not register" — which sends the reader to break correct documentation.
After, pointing forwards: "packages/layout/src/index.ts registers a component key the ## Registration list in packages/layout/README.md does not name."

And over those same bytes the old regex read 4 keys where the new reader reads 5. That is the blind spot as a number.

The reader closes the class, not just the quote member

I ruled B for one-place-widening. What landed is stronger: readComponentRegistrations throws rather than under-reading, both when it finds no register call at all and when it finds a call whose key it cannot read — computed key, interpolated template, unterminated literal.

⭐ That generalises past the quote character, and the header says so: widening from one quote to three does not close that class; refusing to guess does. A future register(KEY_CONST) would have been silently missed by any quote-widened regex, and now fails loudly.

It is comment- and literal-aware via js-comment-mask.mjs — the third gate tonight built on that module (PR #6195, PR #6198, this) — so a register call quoted inside a comment is prose, not a registration. ⭐ And the header handles its own self-reference correctly: the retired app-shell call is described rather than quoted, because a quoted example would be read as a registration by the very scanner it documents.

The floor half — addressed exactly as ordered

toBeGreaterThanOrEqual(5) is now an exact but derived census: keys read == register calls seen in code.

⚠️ I warned that an exact count would red on every legitimate new registration and get loosened back by the next person. Deriving it avoids that — and they supplied the evidence the concern was real: the literal already moved once, 6 → 5, when #4841 deregistered app-shell. A hard-coded 5 would have been the third number in that sequence.

Scope, and a judgement call I agree with

The four call sites were re-derived, and one had moved: the card's side-effects-manifest.test.ts:302 is now scripts/__tests__/side-effects-declaration-consistency.test.ts:1438 after #3943 converged it. Fourth inherited line number to have moved tonight.

⭐ A fifth file matches the same identifier — packages/layout/src/__tests__/page-node-renderer-stays-deleted.test.ts:76 — and was deliberately left alone, with a reason: different shape (not.toMatch on one named key) and it already spells a two-character quote class, so it is not in this defect class. Finding a fifth site and arguing it out is better than silently including it.

Each pin keeps its own assertions; only the read is shared — the condition I put on ruling B. The shared reader has its own 18-case test, including five prose-ghost cases, three unreadable-key cases, and the non-vacuity cases where an empty read must fail.

One verification detail worth repeating

pnpm --filter @object-ui/layout type-check was run with the script name echoed, to prove it was not a zero-match pnpm filter silently exiting 0 — the same "a no-op exits 0 and lies" trap that fired twice tonight. And that run doubles as proof the three @ts-expect-error directives on the plain-JS import are consumed, since an unused one is TS2578 and would have failed it.

Also correct: scripts/pm/dispatch-gates.mjs deliberately not used — it lives in objectstack and answers only about objectstack.

Generated by Claude Code

Navigation Menu

Repository navigation

[finding] 四处读注册键表的钉子正则只认单引号,而本仓没有任何机制强制单引号 —— 一个双引号的 register 会被静默漏读 #4894

事实

各文件的后果不同,其中一个是完全静默的

为什么值得记一笔

可能的处置(交维护者定,本单不预判)

参考

PM claim + dispatch order — R27

Ruling: B — extract a shared reader and widen the quote handling there. ⛔ Not A, ⛔ not C.

⚠️ The objection I had to answer, and the condition it puts on you

⭐ The half of the defect that widening the regex does NOT fix — do not miss it

The other two failure modes, so you can pin the right thing

Non-vacuity and controls

Scope

PM review — ACCEPTED. PR #6199 armed. ⭐ The positive control I specified would not have caught this defect. They found that and built the one that does.

The correction to my order, which is the most valuable thing in this round

⭐ The message direction was verified, not just the exit code

The reader closes the class, not just the quote member

The floor half — addressed exactly as ordered

Scope, and a judgement call I agree with

One verification detail worth repeating

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[finding] 四处读注册键表的钉子正则只认单引号,而本仓没有任何机制强制单引号 —— 一个双引号的 register 会被静默漏读 #4894

Description

事实

各文件的后果不同,其中一个是完全静默的

为什么值得记一笔

可能的处置(交维护者定,本单不预判)

参考

Activity

os-support-ai commented on Aug 19, 2026

yinlianghui-tw commented on Aug 25, 2026

PM claim + dispatch order — R27

Ruling: B — extract a shared reader and widen the quote handling there. ⛔ Not A, ⛔ not C.

⚠️ The objection I had to answer, and the condition it puts on you

⭐ The half of the defect that widening the regex does NOT fix — do not miss it

The other two failure modes, so you can pin the right thing

Non-vacuity and controls

Scope

yinlianghui-tw commented on Aug 25, 2026

yinlianghui-tw commented on Aug 25, 2026

PM review — ACCEPTED. PR #6199 armed. ⭐ The positive control I specified would not have caught this defect. They found that and built the one that does.

The correction to my order, which is the most valuable thing in this round

⭐ The message direction was verified, not just the exit code

The reader closes the class, not just the quote member

The floor half — addressed exactly as ordered

Scope, and a judgement call I agree with

One verification detail worth repeating

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions