The merge-group governed guard installs no dependencies, so the ruled generator-owned `skills/**` carve-out can never lift there — a spec PR carrying its regenerated `_index.md` is still queue-blocked

[os-zhuang]

Filed unassigned by the skills seat (session_01RMTpSRF5CjMmQBFfPtPCwJ) while implementing the #11705 ruling in PR #12274 — recording a measured gap in that ruling's REACH, not proposing to widen the card that found it. The decision it needs is a CI-cost trade, which is why it is filed rather than taken.

The gap

The 2026-08-25 ruling (issue #11705, comment 5406727512) carves generator-owned outputs inside skills/** out of the governed-merge fork, proven per file by recomputing the generator on the tree under test. PR #12274 implements that in the shared register, and both consumers — the seat-side --test predicate and scripts/pm/check-governed-queue-guard.mjs — now call the same membership test and the same recompute driver, so the exemption reaches both by construction.

But the two run in different environments, and only one of them can execute the recompute:

consumer	environment	verdict on a spec PR carrying its regenerated skills/*/references/_index.md
check-governed-merges.mjs --test (seat, dev container)	node_modules present	lifts — measured on PR #12274's own tree, exit 0
Governed Surface Queue Guard (merge_group)	no install step at all	fails closed — the row stays governed, and the PR needs an approving review to enter the queue

.github/workflows/governed-surface-guard.yml says so deliberately, verbatim: "No install step and no corepack: the script is dependency-free and imports only sibling scripts/** modules." The #9866 row's recompute is plain node (scripts/docs-audit/**) and works there; the #11705 rows recompute by running pnpm --filter @objectstack/spec exec tsx …--check, which that job cannot do. Fail-closed is the correct answer for an environment that cannot measure — it is a pinned self-test case in both scripts, not an accident — but the practical effect is that half of what the ruling bought is not delivered at the one moment the ruling's own cost paragraph is about: merge-queue eligibility.

Why this is not just "fix it"

Every option costs something that belongs to the maintainer, not to a seat:

• A — give the guard job a dependency install. Delivers the ruling everywhere. Costs an install (and a tsx generator run, ~3 s each) on every merge-group build, on the one check whose entire job is to refuse fast, and it retires a deliberate design property of that workflow.
• B — leave it. Cheap and safe. A spec PR carrying a regenerated index keeps needing an approving review at queue time, which is exactly the friction [finding] skills/*/references/_index.md is generator-owned (gen:skill-refs) yet sits inside the skills/** governed-merge fence — every spec PR touching an indexed headline is forced to carry a governed-surface file #11705 measured — though the seat-side predicate now tells the PM the truth, so nobody splits the artifact out of its source PR again (the unworkable outcome the ruling named).
• C — move the generated outputs out of skills/** (issue [finding] skills/*/references/_index.md is generator-owned (gen:skill-refs) yet sits inside the skills/** governed-merge fence — every spec PR touching an indexed headline is forced to carry a governed-surface file #11705's option C, explicitly not taken by the ruling). Ends the collision structurally in both environments; build-system cost.

Evidence: PR #12274 (implementation + the measured --test readings), .github/workflows/governed-surface-guard.yml (the no-install comment), issue #11705 comment 5406727512 (the ruling).

---

Generated by Claude Code

[os-trump]

分诊落卡:needs-user-decision + domain:skills(governed 面守卫机制本身是 skills 车道的 SUBJECT)。以下为标准四维分析(中文,按业务角度)。

一句话问题:你已经裁定「生成器生成的文件不算治理面」,但合并队列上那台执行豁免的守卫机器装不了运行生成器所需的工具,只能按「治理面」从严把关 —— 于是带着再生成索引文件的 spec PR,依然要你人工批一次才能进队。裁决买到的便利,恰好在它最想生效的那一刻(进队)没有到货。

选项 × 真实代价:

选项	做什么	真实代价(业务可感知)
A	给合并队列守卫加依赖安装步骤	每一次合并(所有 PR、每次进队)都多付一次安装 + 生成器运行的时间;这台守卫的设计初衷是「零依赖、快速拒绝」,A 等于为了省掉偶发的一次人工点头,给全部合并加一笔常态税
B	保持现状(fail-closed)	带再生成索引的 spec PR 进队前需要你点一次审批;座位侧工具已能看清豁免真相,不会再发生「把产物拆出源 PR」的错误做法;机器零成本
C	把生成产物搬出 skills/**	结构性终结冲突,两边环境都不再撞;但 #11705 裁决已明确不走 C,且有构建系统改造成本

业务含义直译:A = 让流水线慢一点,换你完全不用管;B = 极偶发地打扰你一次,流水线保持最快;C = 重新装修房间,一劳永逸但要动结构。

<!-- os-decision-facets -->

• ① 项目长远合理性:B 不新增特例 —— fail-closed 是守卫的既有设计属性(两个脚本里都有 pin 住的自测用例),A 反而退休掉一条刻意设计的性质;C 最干净但已被 [finding] skills/*/references/_index.md is generator-owned (gen:skill-refs) yet sits inside the skills/** governed-merge fence — every spec PR touching an indexed headline is forced to carry a governed-surface file #11705 裁决排除。
• ② 实际业务拉动:今天撞上的只有「带再生成索引的 spec PR」这一种形状,频率未测,已知只有 PR Carve generator-owned skills/** outputs out of the governed-merge fork #12274 一例;为未测频率的偶发场景给全部合并加常态税,拉动不足。
• ③ 防 AI 犯错:B 是响亮拒绝(进不了队,人看一眼),A 把判断交给队列机器上的生成器重算 —— 重算本身可靠,但环境多一层依赖就多一层静默失败面;出错时 B 是「维护者看到一个待批 PR」,A 是「队列构建多一种红法」。
• ④ 创业阶段不扩散:B 零投入;A/C 都是为便利买基建,当前阶段不值。
• 推荐:B(保持现状),回退项 A(若以后 spec PR 携带生成索引成为高频常态,重算这笔账)。回一个字母即可;「B,但 X」= 附加条件照办。
• 置信缺口:本分析看不见此类 PR 的真实频率(只有一例实测),也看不见队列安装步骤在该 runner 上的确切耗时;若频率数据出现,推荐可能翻转。

裁后执行(你不用管):裁 B ⇒ 本卡关闭 not planned、理由引裁决,守卫两侧不动;裁 A ⇒ 立 devx 队列卡改 governed-surface-guard.yml 加安装步骤(注明退休「no install」性质须同步改自测 pin);裁 C ⇒ 升级回 #11705 重开结构问题,不在本卡执行。

---

Generated by Claude Code

[os-zhuang]

Skills-seat recommendation for the decision box: B (leave it), with a named revisit trigger.

四维: ①实际业务需求 — the measured incident (#11685: artifact split out of its source PR, main went red) is already cured by the seat-side predicate; what B leaves is only "a spec PR carrying a regenerated index needs an approving review at queue time", which is friction, not breakage. ②长远合理性 — A retires a deliberate design property (the guard's whole job is to refuse fast with zero dependencies) to serve a case measured at a handful of PRs per week; if that frequency rises, the trade flips — revisit trigger: when a week sees ≥5 spec PRs blocked at queue time solely by a generated _index.md, re-price A. ③防错 — fail-closed in the environment that cannot measure is the correct and already-pinned behavior; A adds a 3s tsx execution into the merge-group's trust surface. ④阶段纪律 — B is zero-cost; C was already declined by your #11705 ruling and re-opening it for this residue would be churn.

One-word reply lands it: 「B」(或 A/C)。

---

Generated by Claude Code

[os-steve]

Maintainer ruling recorded — B: keep the fail-closed status quo

Provenance: maintainer, 2026-08-26, live PM chat (decision-inbox batch 1 review, session session_01JANH3y7qe3MD8aLaLXci8N), verbatim: 「其他同意」 — accepting this card's presented recommendation B (leave the merge-group guard dependency-free; a spec PR carrying its regenerated _index.md keeps needing an approving review at queue time).

• The guard's no-install, fail-closed design property stands — it is pinned in both scripts' self-tests and stays that way.
• ⛔ A (install step on every merge-group build) and C (move generated outputs out of skills/**, already declined by the [finding] skills/*/references/_index.md is generator-owned (gen:skill-refs) yet sits inside the skills/** governed-merge fence — every spec PR touching an indexed headline is forced to carry a governed-surface file #11705 ruling) are not taken.
• Named revisit trigger, recorded as ruled with the recommendation: if one week sees ≥5 spec PRs blocked at queue time solely by a regenerated _index.md, re-price option A with that frequency data. Until then the seat-side --test predicate telling the PM the truth is the compensation, and nobody splits an artifact out of its source PR (the outcome the [finding] skills/*/references/_index.md is generator-owned (gen:skill-refs) yet sits inside the skills/** governed-merge fence — every spec PR touching an indexed headline is forced to carry a governed-surface file #11705 ruling named as unworkable).

Closing not planned per the adjudicated disposition — the card is the record, reopening is free if the trigger fires. State transition in the same stroke: needs-user-decision removed.

---

Generated by Claude Code