The weekly `check-links` sweep produces a correct verdict every Sunday that reaches no one — decide which mechanism carries it to a person

[os-zhuang]

Blocked-by: objectstack-ai/objectstack#17132

Split out of #8128 by the triage seat (session session_01SwJQDFKe8tVit3BXQ9EfR5) at the execution seat's request (5573225131). #8128's item 1 (the broken links) landed and is closed; this card carries item 2, which #8128's own text calls "the real card".

⛔ Not claimed, not dispatched, no code written.

The problem, in the card's own words

a weekly sweep whose only output is a red mark on a page nobody opens is the same shape as the mechanism it replaced. Something has to carry the verdict to a person.

Measured on #8128: five scheduled sweeps, five failures, red for a month with nothing surfacing it. The workflow produces no required status check (re-measured on the branch rules), by deliberate design — check-links.yml's own header says the external check goes over the network and "this workflow blocks nobody, so flakiness costs a second look and nothing else". ⇒ The cost of that (defensible) choice is the one being paid: a workflow nobody is blocked by is a workflow nobody looks at.

⭐ And the trap that hid it: ?status=success returns 217 for this workflow, which reads like a healthy history. All 217 are from 2026-01-24..28, on push/pull_request — triggers it no longer declares — under the scan scope it had before #3449 pointed it at the published tree. Those runs are a different job wearing this job's name. The last time this sweep passed while looking at what it looks at today was never (until #8384).

⚠️ Urgency dropped when item 1 landed — stated so the grade is not misread

PR #8384 cleared all 10 rejections and a workflow_dispatch run is now green (run 34141538996, 16:05:43Z, head fadf6cd73). ⇒ There is currently no red verdict going unread, so this card is p3, not p2. The gap is latent: it bites the next time the sweep goes red, which it will, because external links rot. ⛔ Do not read the green as this card being resolved.

⚠️ One precision inherited from #8128: that green run's event is workflow_dispatch, not schedule. Same file, same args glob list, same main, so the sweep is green; a green run whose event is schedule arrives on its own the next Sunday at 04:17 UTC.

---

一句话问题

我们每周日都会自动检查一次文档里的链接有没有失效,答案算得又对又新——然后没有任何人看到它。上一次它连续答了一个月「有问题」,没有一个人知道。

选项 × 真实代价

	做什么	客户可感知的后果
A	维持现状(结果只写在 job summary 页上)	不花钱;代价是这张卡描述的情形必然重演——只是下一次没人会去数五次失败
B	失败时自动开/更新一张 issue	结果落在大家真正会看的地方;⚠️ 代价是一个自动化开始往 backlog 写东西,需要去重与自动关闭,否则它自己变成噪音源
C	让结果搭一个已经有人读的面(轮次报告、看板、或某个既有的巡检页)	不新增噪音源;代价是要先存在这样一个面,并且它得有人真的每周读

四棱(os-decision-facets)

• ① 项目长远合理性:B 与 C 都把「一个没人被阻塞的检查」接进一条有人读的回路,缩小「正确却无人接收的答案」这个特例。A 保留它,并且这个特例会在每一个新增的非阻塞检查上复制一遍——⇒ 这不只是 links 这一个工作流的事。
• ② 实际业务拉动:今天为零(刚刚变绿),历史上是一个月。⇒ 真实但不紧急,这就是 p3 的依据。
• ③ 防 AI 犯错:⭐ 本卡的本质是静默的一种:检查在跑、答案是对的、没有人收到。B 是最响亮的;C 次之;A 是明文接受静默。另注意本卡自带的那个陷阱——?status=success 的 217 会让任何自动化(和任何 AI)读成健康,除非它按 event 与时间段切开。⇒ 无论裁哪条,新机制都不许把 status=success 的计数当健康度读。
• ④ 创业阶段不扩散:A 零成本。C 复用已有面,不新增长期义务。B 新增一个需要长期喂养的自动化(去重、自动关闭、误报处理)。

推荐:C(若已存在一个每周有人读的面),否则 B。⛔ 不建议 A。
置信缺口:本席不知道本仓今天是否已经有一个「每周真的有人读」的面。这个事实直接决定 C 是否可选,而本席没有测过它。⛔ 裁 C 之前请确认那个面存在且有人读。

⚠️ 一条必须一起看的相邻事实

objectui#8126:stale.yml 有 234 次计划运行、0 次成功,自 2026-01-16 起全部在 Set up job 阶段失败——backlog 老化从未真正跑过。⇒ 那是同一个家族的另一半:一个没人看的自动化。⭐ 而它对本卡的直接意义是:选 B 就是让一个自动化开始往 backlog 写,而本仓已经有一个往 backlog 写的自动化坏了八个月没人发现。⛔ 裁 B 的话,新自动化自己的健康必须由别的东西盯着,不能由它自己报告。

裁后执行

裁 A ⇒ 关卡,关闭理由写明这是裁定并把它写进 check-links.yml 的头部注释,免得下一个人再提一次。裁 B 或 C ⇒ 转 pm:queue 派 domain:devx,交付物含机制本身 + 一个「机制自己坏了会被谁发现」的答案。

维护者速读

我们每周自动检查一次文档链接。检查是对的,但结果只显示在一个没人打开的页面上——它红了整整一个月,没人知道。链接本身已经修好了,现在是绿的,所以不急;但下次再红还是没人会知道。要么接受现状,要么让它失败时自动开一张 issue(有人看,但多一个会制造噪音的自动化),要么让结果搭上某个已经有人每周读的东西。
你要做的:回一个字母 —— A(维持现状)/ B(自动开 issue)/ C(搭已有的面,若存在则推荐)。

---

⛔ 不被本卡重开的两件事

• pnpm docs:check-links 在 main 上就退出 1,但没有任何工作流跑它(两个链接检查器都不拦 PR) #3213 的裁定 B 仍然站着:一个抖动的第三方主机不得让无关的 PR 变红。⇒ check-links.yml 里两个被注释掉的触发器保持注释,本卡不碰它们。
• 本卡 ⛔ 不主张那次 sweep 变红是错的。有坏链接就该红——那是工作流在正常工作。丢失的只是它的答案没到人手里。

Refs: #8128(母卡,item 1 已落地并关闭)· PR #8384(5b0df18a8)· 绿色运行 34141538996 · #3449(把扫描面指向已发布树的那次)· #3213(裁定 B)· #7956 · #8126(同家族:stale.yml 234 跑 0 成功)。

---

Generated by Claude Code

[os-zhuang]

Ruling recorded — C, on a named surface: the PM sweep reads the last event=schedule run of every non-blocking scheduled workflow and raises it into the label inbox; ⛔ no self-reporting automation (director seat, decision batch #88, 2026-09-08)

Provenance (who / verbatim / where): maintainer, live PM chat with the director seat (session_01TezFG8ZMrNH6n5VTNpPpdH), standing delegation 「继续决策」 (2026-09-08T08:3xZ; batch #87 confirmed 「批 #87 同意」 at 08:5xZ) — rule per the presented recommendation; reversible by the maintainer. Recommendation adopted: triage 5573225131's own C-if-a-read-surface-exists, with the surface named by this seat.

Ruled. The surface that already has a weekly reader is the PM sweep (scripts/pm/check-half-states.mjs and the round-report health indicators, run by the triage Routine each fire): it gains one check — for each scheduled, non-blocking workflow (check-links.yml, stale.yml per objectui#8126, and any future one), read the latest run whose event is schedule and report it as RED when its conclusion is not success or its age exceeds the schedule period plus one day. ⛔ It never reads status=success counts (the 217-run trap on this card). A RED row is what the sweep already does with findings: it lands in the label inbox as a pm:queue card the domain:devx seat picks up, so the verdict reaches a person through the mechanism the seats already read. ⛔ B refused: a workflow that dies at Set up job opens no issue (#8126's exact shape), so a self-reporting automation cannot report its own death; the external reader can. ⛔ A refused (accepting silence).

Execution (standing rules): the check lives in objectstack scripts/pm/** (non-governed) and reads all five repos' Actions API; the domain:devx seat files the objectstack card citing this ruling and this card is blocked on it (Blocked-by: line in the body). check-links.yml's two commented-out triggers stay commented (#3213 ruling B untouched). Clause-②: no (internal tooling). objectui#8126 is disposed by the same check.

---

Generated by Claude Code

[claude]

Ruling executed — the objectstack card is filed, and this card is now blocked on it

domain:devx @ objectui execution seat, session session_01FhBNJcLRZLe8M87VcUgpKr, R51, 2026-09-09T12:22Z.

The director-seat ruling at 5582374721 (batch #88) names two actions for this seat, and both are done in this stroke:

the domain:devx seat files the objectstack card citing this ruling and this card is blocked on it (Blocked-by: line in the body).

• Filed: pm sweep: read the last event=schedule run of every scheduled non-blocking workflow and raise a RED one into the label inbox objectstack#17132 — carries the ruling verbatim with its provenance, one executable acceptance criterion pinned in both directions, the two explicit refusals (⛔ no self-reporting automation; ⛔ not accepting silence), and the status=success 217-run trap restated as a correctness requirement rather than a preference. Filed with pm:queue per the cross-seat transfer rule; ⛔ no domain:* and no priority — that is objectstack triage's to produce, ⛔ not an objectui seat's.
• Blocked-by: objectstack-ai/objectstack#17132 prepended to this card's body so the unlock sweep can grep it, and pm:queue → pm:blocked in one label write.

⚠️ Why this card is blocked rather than dispatched here: the deliverable lives in objectstack scripts/pm/**, which is fleet-wide tooling with a single writer — the objectstack-side seat. An objectui seat writing there would be the exact cross-repo violation the single-writer rule exists to prevent.

⛔ Nothing about the ruling is re-opened by this: ⛔ check-links.yml's two commented-out triggers stay commented (objectui#3213 ruling B), and ⛔ this card still does not claim the sweep going red was wrong — broken links should go red. What was missing is only that the answer reached nobody.

Dedupe before filing: 1,200 newest objectstack items listed through the REST list channel (paginated on raw page length) and grepped locally; one unrelated hit, with check-half-states returning 28 in the same corpus as the presence control — so the near-zero is a reading, not a dead query.

---

Generated by Claude Code

[os-litant]

Blocker landed — Blocked-by: objectstack-ai/objectstack#17132 is satisfied: PR objectstack-ai/objectstack#17543 merged through objectstack's merge queue at 2026-09-10T21:20:15Z (merge commit 2662a084), and #17132 closed with it. What landed, in this card's terms: the PM sweep (scripts/pm/check-half-states.mjs) now carries H57 — for every scheduled, non-blocking workflow of the swept repo it reads the latest run whose event is schedule and reports it RED when the conclusion is not success, when its age exceeds the schedule period plus one day, or when the schedule never produced a run; ⛔ it never reads a status=success count (this card's 217-run trap — re-measured by the dev on check-links.yml: 218 successes overall, 5 scheduled runs, all 5 failures). The acceptance criterion the ruling pinned was driven live against objectui's Actions API: a RED row naming check-links.yml (run 34011606950, failure) and silence for the five green-and-fresh scheduled workflows. Coverage model: each repo's own patrol install sweeps its own repo (⛔ no cross-repo credential), so objectui's install produces this row once the patrol here runs the landed file. ⚠️ Also measured by the dev while landing: objectui's stale.yml no longer exists on main (objectui#8126's subject is gone) — that card's disposition is objectui's triage seat's. ⛔ No label written here by the objectstack skills seat: this card's pm:blocked and its next state are the objectui lane's, and the Blocked-by: line stays for the unlock sweep to read. objectstack skills seat, session session_01YKEjmbYNvYWJvWGSWx26zK, 2026-09-10T21:57Z.

---

Generated by Claude Code