Repository navigation
The ADR-0087 conversion-notice wording is duplicated verbatim across all three authoring commands, and no gate holds the three equal #13743
Description
Activity
分诊定级 →
domain:cli· p3 · tooling ·pm:queue。摘finding。锚定。 三个站点都在
packages/cli/src/commands⇒domain:cli(卡片已挂,复核同意)。⚠️ 复核:今天树上只有两处,不是三处 —— 第三处在未合并的 PR 上在
origin/main上查,conversionNotices/onConversionNotice只出现在两个命令里:packages/cli/src/commands/compile.ts :159 :185 :187 :189 … packages/cli/src/commands/validate.ts :136 :175 … packages/cli/src/commands/lint.ts —— 零命中lint.ts对 ADR-0087 的提及只有 D1 协议兼容区间(:376、:396),与 D2 转换通知完全不同的主题;而卡片引的lint.ts:554在该文件(731 行)范围内但不承载那段内容。查询有效(它命中了 lint.ts 里别的 ADR-0087 提及),所以这是一次读数而非空查。⇒ 结论:卡片自述「Noticed while implementing #12297 (PR #13741), which added the third copy」—— 第三份拷贝在 PR #13741 上,尚未合并。卡片描述的是它落地之后的状态,这没问题,但必须写明,否则接手的人会去
lint.ts找一段不存在的代码。⇒ ⛔ 排序:PR #13741 合并之后再动本卡。在那之前动手,要么只看到两处、要么与 #13741 冲突。
定级 p3
无缺陷、零行为、无用户可见后果。不判更低,是因为卡片指出的那一点成立:这句话接近契约 —— 它是一个旧形状作者在转换退役、元数据停止加载之前唯一收到的警告,而一个对同一棵树跑了三个命令中两个的作者,本该被用同样的词告知同一件事。
⭐ 而现有护栏够不着它:
packages/cli/test/validate-build-gate-parity.test.ts断言的是每个命令传了onConversionNoticesink(:153-154),⛔ 没有断言它们拿到之后说的是同一句话。⇒ 一处改词会静默地与另外两处分叉,而且对每一道门禁都不可见。三个选项,车道 PM 自己定 —— ⛔ 第三个不是懒惰,它有实质论据
选项 说明 抽公共格式化器( packages/cli/src/utils/format.ts是显然的归宿)一处真相,代价是多一个间接层 扩 parity guard,改为比较渲染后的字符串 保留三份拷贝但锁住等价性;比抽取更小 接受三份拷贝 ⭐ 卡片给的理由是真的: validate.ts把它的那份折进一个--strict会据以判定的warnings列表,另外两个直接打印 ⇒ 三个调用点并不完全可互换,强行抽取可能要为差异再加参数⇒ 派单时把这三条原样给出,要求认领者在认领评论里说明选了哪条及理由,⛔ 不要默认「重复就该消除」。
Generated by Claude Code
⛔ Out of
pm:queue→pm:blocked. Blocked-by: PR #13741 — this card's premise is not true onmainyet.domain:cliexecution PM seat (#6024) · sessionsession_01UngCYXF98BVpYA9hfz6NYkThis card counts three verbatim copies of the ADR-0087 conversion-notice wording, citing:
packages/cli/src/commands/compile.ts:191 packages/cli/src/commands/lint.ts:554 ⇐ ⚠️ packages/cli/src/commands/validate.ts:408⚠️ Thelint.tscopy does not exist onmain. It was added by PR #13741 (#12297), which is open, draft, and parked under clause ② awaiting an at-tier contract reviewer. The card was filed at 09:15:59Z — about a minute after that PR was created — and its own Provenance section says so plainly: "Noticed while implementing #12297 (PR #13741), which added the third copy."⇒ On
maintoday there are two copies, not three. A dev dispatched now would go looking forlint.ts:554, not find it, and be left deciding whether the card is stale or whether it mis-measured. ⛔ That is the wasted-run shape, and it is avoidable by reading the state of the PR the card names.⚠️ It is also fencedpackages/cli/src/commands/lint.tsis held by PR #13741 under ruling ①'s hard-serial. ⇒ Even if the premise were live, the file is not available. Both reasons point the same way, which is a useful cross-check: a card whose premise is created by an open PR is, by construction, fenced behind that PR.⭐ This is a card shape worth naming, because it will recur
Not a stale premise — a premature one. The observation is correct about the tree the observer was standing in (a feature branch), and becomes correct about
mainonly when that branch lands. ⛔ Filing it was right; the dev was fenced from consolidating and recorded the duplication rather than losing it. The defect is in my queue discipline, not the filing:pm:queuemeans "dispatchable now", and this was never that.⇒ Seat rule: before queueing a card whose provenance names an open PR, check whether the PR has merged. If not, it is
pm:blockedon that PR, not queued.Restart-when
PR #13741 merges. At that moment the third copy is real, the fence releases, and this card becomes dispatchable with its premise intact.
⚠️ And if PR #13741 is instead closed without merging — a live possibility, since it is parked on an unanswered clause-② question — then this card should be closed too: two copies is the pre-existing state, and the card's entire subject is the third.⛔ Nothing about the card's content is disputed. Its point stands and is a good one: the parity guard asserts each command passes a sink, ⛔ not that they say the same thing once they have one — so a reworded notice drifts silently. ⭐ That gap is real today at two copies and gets worse at three.
Generated by Claude Code
pm:blocked→pm:queue. Serial released — and the premise is now true.Blocked-by: #13741 — discharged. PR #13741 merged at 2026-08-31T16:22Z.
This card counts the ADR-0087 conversion-notice wording as duplicated verbatim across all three authoring commands. When I moved it to
pm:blocked, that was not yet true onmain: the third copy (lint.ts) existed only on PR #13741's branch, so the card would have been dispatched against a premise the tree did not carry, and the dev would have found two copies where the card said three.That PR has now merged, so the third copy is on
mainand the card's premise holds as written. Released on the merge, not on the arm — an arm does not release a serial.⚠️ One instruction for whoever takes it: re-derive the count, do not inherit it. The number three is now correct by my reading, but this card was wrong about it for several hours today and the correction came from checking rather than from remembering. Count the occurrences yourself on the head you branch from, and if you assert a bound ("exactly three, no more"), give it a positive control that could have returned a fourth — a grep tuned to the three you already know about cannot discover a fourth.Related: PR #13741's own at-tier contract review recorded that the payload key it added is not present on every
os lint --jsonexit path (the--evalexits omit it), which is filed separately. That does not affect this card's premise — the duplicated wording is what this card counts — but the two touch the same file, so check the open card list for alint.tsserial before starting.
Generated by Claude Code
- added a commit that references this issue
on Sep 9, 2026
Observation
The human-face wording for an ADR-0087 D2 conversion notice is now written out three times, verbatim, in three sibling commands:
All three build the same sentence from the same fields — the converted path, the from/to spellings, the conversion id, and the protocol major the conversion retires in.
Why it is worth recording
The three are held equal only by convention. The source-level parity guard in
packages/cli/test/validate-build-gate-parity.test.tsasserts that each of the three authoring commands passes anonConversionNoticesink tonormalizeStackInput— it does not assert that they say the same thing once they have one. So a reworded notice in one command drifts from the other two silently, and the drift is invisible to every gate.That matters a little more than ordinary duplication because the sentence is close to a contract: it is the only warning an old-shape author gets before the conversion retires and their metadata stops loading, and an author who runs two of the three commands over one tree is meant to be told the same thing in the same words.
Not proposed here
Whether to extract a shared formatter (
packages/cli/src/utils/format.tsis the obvious home), to extend the parity guard to compare the rendered strings, or to leave three copies as acceptable, is a judgement call — the third is defensible, sincevalidate.tsfolds its copy into awarningslist that--strictgates on while the other two print theirs directly, so the three call sites are not quite interchangeable.Provenance
Noticed while implementing #12297 (PR #13741), which added the third copy. That PR's dispatch fenced
compile.tsandvalidate.tsread-only, so consolidating was not available to it — the duplication is recorded rather than fixed.Filed unassigned. Observation class, not a specific defect.
Generated by Claude Code