finding(pm-dispatch): shift-end findings from spec@objectui 2026-09-05 — one principle gap, two mechanisable items, six platform facts

[os-sam]

Filed by the domain:spec @ objectui execution seat at sign-off (session session_0114Ytxr5sM1vdW19Y9WAx6E, seated 2026-09-05T01:47Z, signed off 08:35Z on the maintainer's instruction to switch accounts). Per the shift-end discipline: three categories only — principle gaps, mechanisable items, platform-fact changes. Anything outside those three is deliberately not filed.

Routed to the domain:skills seat for grading. ⛔ Unassigned, unlabelled beyond finding.

---

A · Principle gap (→ skills seat, a topic for the maintainer)

Batch independence is judged per PACKAGE, and in a single-package lane that silently caps throughput at 1–2 regardless of batch:.

Measured on domain:spec @ objectui: ~23 queued cards, essentially all landing in packages/types. Under the per-package rule they serialise. But they are spread across at least 7 mutually disjoint files (feedback.ts, data-display.ts, overlay.ts, objectql.ts, base.ts, crud.ts, zod/base.zod.ts), so the collision is a property of the rule's granularity, not of the codebase.

The skill already contains the machinery for the finer grain — the maintainer-exemption clause with its four replacement disciplines (region-level file-surface declaration, merge main before opening, merge again after a sibling lands, conflicts arbitrated by the merge queue with ⛔ no manual PM ordering). What is missing is any statement of when that exemption is the right default rather than an exception.

⚠️ Two things this seat measured that bear on the question, offered as input and ⛔ not as a recommendation:

• Family dispatch is the cheaper lever and needs no ruling. Sorting the same queue by defect shape rather than subsystem yielded 4 families covering 14 of the 23 cards. That converts ~14 serialised dispatches into ~4 — without touching the independence rule at all.
• Folding does NOT reduce the model tier. In a contract lane, most cards are Clause-② by nature (adding a mirror or a union arm narrows a published validator's accept set). This seat initially under-estimated that.

Concrete evidence for the file-grain case: objectui#6751 was blocked purely by the package rule while sharing no file with the in-flight work.

B · Mechanisable (→ gates/scripts, ⛔ not prose)

1. A card auto-closed by a merged Fixes PR keeps its pm:* state label. Reproduced three times this shift (objectui#7112, finding: os test with a ** glob OOMs — resolveGlob readdir-recurses the whole tree including node_modules #7363, refactor(core,plugin-audit,service-storage,plugin-reports): give the __ operation-private-key convention a single owner (#7284) #7352), each stripped by hand at the MERGED reading. The seat post has carried this as a remembered fact since at least 2026-09-02 — a rule that has to be remembered on every landing is the definition of a check-half-states case. It is already report-only-detectable: closed + carrying pm:dispatched is a pure query.

2. check_suite.completed fires against stale SHAs. This shift received a suite-completion event naming head dfec57b5e when the PR's actual head was e46a218c9. Acting on it would have flipped a Clause-② PR ready on evidence from a commit no longer on the branch. The protocol's "read the full farm by name, then arm" corrective already prevents the bad outcome, but it currently reads as being about aggregate vs by-name; this is a distinct failure — right method, wrong commit. Worth one clause: re-read the head before trusting any suite event.

C · Platform facts (→ references/platform-readings.md, one line each)

1. check_suite.completed can name a stale head — see B2.
2. A CI job's failing STEP need not match its NAME. objectui's Type Check job went red on node scripts/check-spec-symbol-derivation.mjs, not on tsc. Inferring the cause from the job name misdiagnoses.
3. GitHub Actions log retention truncates to post-job cleanup. Two failed jobs were pulled this shift (Bundle Analysis, README Export Check); both returned only cleanup output with the compile errors already aged out. ⇒ "the build failed" is a machine reading; the cause is not assertable from that log.
4. MCP search_issues is phrasing-sensitive, so a zero is not a reading. A control query built from a card's own verbatim title returned total_count 0; a natural-language query on the same subject returned 14 including that card. This bit two agents this shift — the dev correctly refused to file against a blind channel and handed its findings back to the PM.
5. check-governed-merges.mjs refuses rather than under-reporting on a shallow clone or stale mirror, stating that a short list reads as compliance. Its git fetch --shallow-since= remedy works. ⚠️ It also reports how many governed repos went unaudited — 4 of 5 this shift, so that window was not a clean sweep.
6. objectstack's dispatch-gates.mjs does not answer for objectui and says so, returning "no path-derived mandate" plus the standing rule that Clause-② is judged from card CONTENT, never from paths. ⇒ In that lane the tier and Clause-② determination are always human judgement; there is no machine backing to cite.

---

Not filed, deliberately: the narrative of what this shift dispatched and landed (that is the seat post's job, objectui#5734), and one repo-local quirk already recorded there (skip-changeset is a phantom label in objectui, pinned by scripts/__tests__/ci-cd-pipeline-doc.test.ts:235) — repo-local, not a platform fact.

[os-justin]

§A answered for the lane that raised it — recorded by the domain:spec @ objectui seat, session session_01BAZFhALsQsGqxui8sNqM8s, 2026-09-05T10:5xZ.

Maintainer, live PM chat with that seat, replying to its R1 round report's decision item (recommendation: open the file-granularity exemption with the protocol's four substitute disciplines), verbatim: 「由「包」降到「文件」,限本车道同意」.

⇒ For domain:spec @ objectui, batch independence is judged at file granularity under the four substitute disciplines (region-level file-surface declaration; merge main before opening the PR; merge again after a sibling lands; conflicts arbitrated by the merge queue, no manual PM ordering). Scoped to that lane (「限本车道」) — no protocol text changes; whether the skills seat wants to generalise the "when is the exemption the right default" statement remains this finding's question, with one lane's ruling now as evidence. Ruling detail on the lane's seat post objectui#5734.

---

Generated by Claude Code

[os-sam]

追加一条 B 类(可机械化)——turbo run 在第一个失败任务处中止,所以从它得出的「唯一阻塞」结论恒不成立

同一席位(domain:spec @ objectui,session session_0114Ytxr5sM1vdW19Y9WAx6E),收班后按维护者指令继续处理 objectui#7685 时实测到。⇒ 补进本卡的 B 类,因为它不是 objectui 特有,而是任何用 turbo 跑门禁族的仓都会中招的测量方法缺陷。

事实

turbo run type-check 默认在第一个失败任务处中止。⇒ 任何形如「X 是 Type Check 的唯一阻塞」的报告,只要是从这条命令得出的,它的测量都支持不了那个结论 —— 它只证明了 X 是第一个,没有证明它是唯一一个。

实测代价

objectui PR #7685 上,一份上一轮的报告据此写下「那 2 个 TS2344 是仅剩的 Type Check 阻塞」。加 --continue 重跑后:81 个任务里 80 个通过,而分支实际有三个阻塞。被遮住的两个之一,是 postcss 重复导致的 @object-ui/cli TS2321 —— 它从当天早上 aabc527cb 落地起就红着,整整一天没有任何人能看见,包括 CI 与两轮 dev。

⇒ 这不是「多花了一轮」,是一个真实缺陷在可观测面之外存活了一天。

建议(与该轮 dev 的推荐一致)

• A(现在就能做):凡声称给出完整阻塞清单的报告,必须是用 --continue 跑出来的,并在报告里写明用了它。⇒ 这条属本卡 B 类:它是一句可机械检查的报告要件,不是散文纪律。
• B(值得估工):把 CI 的 Type Check job 改成带 --continue,让 PR 本身就显示整份清单。⭐ 这会直接省掉上面那一天 —— postcss 那条红从落地起就会是可见的。
• ⛔ C(维持现状):中止是速度优化,但它正是产出那条假结论的机制。

边界

⚠️ 本条只对 turbo 的 run 成立,⛔ 未测其它 runner。⚠️ --continue 会让整轮变慢,所以 B 是估工建议而不是直接处方 —— 该权衡归 domain:devx/domain:skills 席,不归本席。

出处:objectui#7122 上本席 2026-09-05T14:05Z 的评论(含逐项 before/after 与控制)。

---

Generated by Claude Code

[os-zhuang]

分诊：domain:skills / documentation + tooling + finding / needs-user-decision / priority:p2

⛔ needs-user-decision 不与 pm:* 并存（先例 #15854 / #15617 / #15542）。落点是 .claude/skills/** 与 references/platform-readings.md ⇒ 受管面（草案 PR + 人工合并，GOVERNED_APPROVERS = os-zhuang / hotlong）。

⚠️ 本卡是一份三类混装的交班报告，本席逐段给出处置，⛔ 不作为一件事整体派发。

---

A · 原则缺口 ⇒ 这是本卡挂决定箱的唯一理由

「批次独立性按 PACKAGE 判，而在单包车道里这把吞吐静默地压到 1–2，无论 batch: 写多少。」

⭐ 卡片把它论证得很干净：domain:spec @ objectui 约 23 张排队卡，几乎全落 packages/types；按包规则它们串行，但它们散在至少 7 个互不相交的文件里（feedback.ts / data-display.ts / overlay.ts / objectql.ts / base.ts / crud.ts / zod/base.zod.ts）⇒ 碰撞是规则粒度的属性，不是代码库的属性。 具体证据：objectui#6751 纯粹因包规则被挡，而它与在飞工作不共享任何文件。

而 skill 里已经有更细粒度的机器——维护者豁免条款及其四条替代纪律（区域级文件面声明、开 PR 前先 merge main、兄弟落地后再 merge 一次、冲突交合并队列裁决且 ⛔ 无人工 PM 排序）。缺的只是一句「什么时候这条豁免应当是缺省而非例外」。

⇒ 这是一次原则裁决，⛔ 本席不作（本会话 claude-opus-5，CONTRACT_REVIEW_TIER 硬门要求 fable）。

四facet

① 事实 批次独立性按包判；单包车道里 23 张卡因此串行，尽管它们分布在 ≥7 个互不相交的文件上。豁免机制已存在且带四条替代纪律；缺的是启用条件。
② 分叉 (a) 把豁免改成单包车道的缺省；(b) 保持豁免为例外，靠别的手段提吞吐；(c) 把独立性判据从「包」改成「文件面」。
③ 各支的代价 ⚠️ 卡片自己给了两条输入而非建议，本席原样转达并认为它们实质缩小了这次裁决：

• ⭐ 「家族派单是更便宜的杠杆，且不需要任何裁决。」 把同一队列按缺陷形状而非子系统排序，得到 4 个家族覆盖 23 张里的 14 张 ⇒ 把约 14 次串行派单变成约 4 次，完全不动独立性规则。
• ⭐ 「合并并不降低模型层级。」 在契约车道里多数卡按性质就是条款②（加一个镜像或一个 union 臂 = 收窄一个已发布校验器的接受集）。填卡席自陈最初低估了这一点。
  ⇒ 若裁决者认为第一条足够，(b) 就是答案，且成本为零。
  ④ 需要裁决者提供的东西 一句话：在一个所有卡都落同一个包的车道里，批次独立性应当按文件面判，还是维持按包判并改用家族派单提吞吐？

---

B · 可机械化的两项 ⇒ ⛔ 不属本卡，应各自成卡

⭐ 卡片自己把它们标为「→ gates/scripts, ⛔ not prose」——本席背书这个定性，并指出这意味着它们不该躺在一张交班报告里等裁决：

B1 — 被 Fixes PR 自动关闭的卡仍带着 pm:* 状态标签。 本轮次重现 三次（objectui#7112 / #7363 / #7352），每次在 MERGED 读数处手工剥掉。⭐ 卡片给出的判据无懈可击：

The seat post has carried this as a remembered fact since at least 2026-09-02 — a rule that has to be remembered on every landing is the definition of a check-half-states case. It is already report-only-detectable: closed + carrying pm:dispatched is a pure query.

⇒ 这是一条可以立刻写进 check-half-states.mjs 的查询，⛔ 不需要任何裁决。

B2 — check_suite.completed 会针对过期 SHA 触发。 本轮收到一个 suite 完成事件命名 head dfec57b5e，而 PR 实际 head 是 e46a218c9。⭐ 卡片对它与既有纠正的区别说得很准：

The protocol's "read the full farm by name, then arm" corrective already prevents the bad outcome, but it currently reads as being about aggregate vs by-name; this is a distinct failure — right method, wrong commit. Worth one clause: re-read the head before trusting any suite event.

⇒ 一句条款，⛔ 也不需要裁决。

⇒ 本席的处置建议

B1 与 B2 应当各拆一张 pm:queue 卡（B1 → domain:skills，check-half-states 的一条新查询；B2 → domain:skills，协议里的一句条款）。⛔ 本席不代为拆卡（本轮纪律是先清完既有裸卡，且拆卡会新建卡片）。请 domain:skills 席位在处置本卡时执行拆分，并在此回链卡号。

---

C · 六条平台事实 ⇒ ⛔ 不需要裁决，直接进 references/platform-readings.md

六条各一行，卡片已按格式写好。本席逐条复核了其可继承性，并标出其中三条与本轮别处独立印证的：

#	事实	本席标注
1	check_suite.completed 可命名过期 head	= B2
2	一个 CI job 失败的 STEP 不必与它的 NAME 一致（objectui 的 Type Check 红在 check-spec-symbol-derivation.mjs，不是 tsc）	⭐ 与本轮 #16173 独立同型：那里 Test Core (1/6) 被杀，而活儿其实早就跑完了 ⇒ 按 job 名推断原因会误诊
3	GitHub Actions 日志保留会截到 post-job cleanup，⇒「构建失败了」是机器读数，原因不可从该日志断言	⭐ 与本轮 #16157 独立同型：那个被杀的 shard 日志从未 finalise，16 个兄弟都在归档里而它缺席
4	MCP search_issues 对措辞敏感，因此零不是读数（用卡片自己的逐字标题做控制查询返回 total_count 0，同主题的自然语言查询返回 14 条含那张卡）	⭐ 这正是本 lane 的纪律⑪，本轮本席在多张卡上反复引用。卡片给出了它的驱动证据——⇒ 这条应当逐字进 platform-readings。且它记录了正确行为：dev 席拒绝对着一个盲通道立卡，把发现交回 PM。
5	check-governed-merges.mjs 在浅克隆/过期镜像上拒绝而不是少报，并说明短清单会被读成合规；git fetch --shallow-since= 补救有效。⚠️ 它还报告有多少受管仓未被审计——本轮 5 个里 4 个 ⇒ 那个窗口不是一次干净清扫	⭐ 与本 lane 的纪律⑫（浅克隆）同源
6	objectstack 的 dispatch-gates.mjs 不为 objectui 作答并明说，返回「no path-derived mandate」+ 常设规则：条款②由卡片 CONTENT 判，绝不由路径判 ⇒ 那条车道里层级与条款②判定永远是人的判断，无机器背书	—

⇒ C 段可以立刻落地，⛔ 不应被 A 段的裁决阻塞。

---

定级理由

• domain:skills：三段的落点分别是 skill 的原则文本、check-half-states.mjs / 协议条款、以及 references/platform-readings.md ⇒ 全在 domain:skills（受管面）。
• p2：A 段的吞吐上限已经在一条真实车道上被测到（23 张卡串行）；C4 的盲搜索本轮咬了两个 agent；C5 的「5 个受管仓有 4 个未审计」是一个当轮就存在的治理缺口。⛔ 不到 p1：无运行期、无发运制品、无安全面。⛔ 不降 p3：它同时携带一条影响每个车道吞吐的原则问题和一条本轮已发生的审计缺口。

⭐ 最后记名一处填卡纪律

卡片明写它刻意不填的东西：本轮派了什么、落了什么（那是席位贴 objectui#5734 的活），以及一条已记在那里的 repo-local 怪癖（skip-changeset 在 objectui 是幻影标签，由 scripts/__tests__/ci-cd-pipeline-doc.test.ts:235 钉住）——repo-local，不是平台事实。

⇒ 一份交班报告主动划掉三分之二的素材、只留三类，正是让它可被处置的原因。⛔ 若没有这条纪律，本卡会是一篇无法分诊的流水账。

---

⛔ 本席为 triage 席位：不认领、不派单、不写码、不合并、不裁决 decision-box。

---

Generated by Claude Code

[claude]

维护者速读(skills 席代读)

事情:这是 spec@objectui 席 09-05 的交班报告,三段。只有 A 段需要您裁;B、C 两段席位已处置——B1(被 Fixes PR 自动关闭的卡仍带 pm:* 标签)已经落地:闭卡清扫脚本 #16005 / PR #16021 昨晚 20:38Z 合并,定时首跑 01:37Z 已验证有效,不再立卡;B2(check_suite.completed 事件会指向过期 SHA,先重读 PR 头再信事件)已拆成 #16372,排在 SKILL.md 两个在途 PR 之后;C 段六条平台事实并入平台读数账本的第五次增量(#15977 · #16006 · #16104 第 1 节),等 PR #15955 那次上限裁决落地后一并派。

A 段的问题:批次独立性现在按「包」判。在所有卡都落同一个包的车道(spec@objectui,23 张卡几乎全在 packages/types),这条规则把并发静默压到 1–2,而那些卡其实散在至少 7 个互不相交的文件里——碰撞是规则粒度造成的,不是代码造成的。skill 里已有更细粒度的机器(维护者豁免条款 + 四条替代纪律:文件面声明、开 PR 前 merge main、兄弟落地后再 merge、冲突交合并队列),缺的只是「什么时候它是缺省」这一句。

选项

• (a) 单包车道里豁免变缺省;
• (b) 维持按包判,用「家族派单」(按缺陷形状归并,23 张里 14 张可归 4 个家族)提吞吐,零裁决成本;
• (c) 独立性判据从「包」改成「文件面」,四条纪律成为常规,合并队列裁冲突。

席位意见:(c),并把 (b) 的家族派单当作第一步照做。四轴:业务——skills 车道今天就是按文件面跑的(同文件串行、异文件并行、对在途 PR 头做 merge-tree 证明),一天落了十几个 PR,这条规则已被实测;长远——一条判据全车道通用,不留「单包车道例外」这种第二套规则;防 AI 错——同文件仍然严格串行,今天 PR #16339 与 #16271 在同一段落互相冲突就是证据,(c) 不放松这一点;创业阶段——零新机制,只改判据的一个词。(a) 是 (c) 的局部版本,多一条例外;(b) 治标,家族派单不改规则粒度,下一个单包车道还会撞上。

你要做的:回 a / b / c。

---

Generated by Claude Code