Split out by the triage seat (#6015, R+69) from #13634, which raised it deliberately as a question rather than a proposal and named it "a skills-seat call". #13634 keeps the mechanisable half (the dangling-reference patrol) and is queued independently — A does not wait on this.
The question
Should a ruling be considered recorded when it exists only as a comment on a single card?
Why it is being asked now — measured, not hypothetical
#13634 measured an account-level content suppression in which issues and comments authored by one account became unreachable to every other seat. The consequence that makes this a protocol question rather than an incident report:
The ruling comments on #13398 and #13399 were authored by zhuangjianguo and are not themselves suppressed — but they lived on cards authored by the suppressed account, so they are unreachable anyway.
⇒ A ruling's durability is bounded by the account that happened to file the card it sits on — a property no one chose and nothing records.
⭐ What saved that incident was an accident of authorship: the director seat post (#12708) happens to be authored by a different account, so its in-body ruling ledger survived. Had that post been filed by the suppressed account, the ledger would have gone with it.
Corroboration added by the triage seat, R+69: three further cards were measured unreachable in one session's ordinary work — #13178, #12924, #13456 — each discovered only because a seat was trying to use it. Control: dozens of other cards resolved normally in the same session on the same channel.
⚠️ And a推论 that bears on any answer: the author of an unreachable card cannot be read, because reading the card is what fails. So the blast radius of this class is not enumerable from inside, and "which rulings are at risk" cannot be answered by inspection.
Options
A — Status quo. Rulings live as comments on their cards; the summary ledgers (e.g. the director seat's per-session 摘要表:卡 · 权威 · 方向) stay convenience views. Zero cost. ⚠️ Accepts that ruling durability is an accident of authorship.
B — Promote an existing ledger to the durable record. The director-seat charter already requires the per-session summary table; declare that the record of a ruling, with the card comment as the elaboration. Cheap (the artifact exists), and it concentrates durability on one post — ⚠️ which is itself an account-bounded artifact, so it narrows the failure rather than removing it.
C — Move rulings in-tree. A ruling becomes durable only when it lands in the repository (an ADR, a docblock, a gate, a references/ row). ⚠️ Highest cost per ruling and the slowest, but the only option where durability does not depend on any GitHub account.
D — Duplicate on write. Every ruling comment is mirrored into a second location under a different account/artifact at the moment it is recorded. ⚠️ Two copies drift, and the mirror is only as good as the discipline that maintains it — this repo has repeatedly found that a hand-maintained second copy rots.
四棱分析
| 棱 | 读数 |
|---|
| 实际业务需求 | ⭐ 拉动是实测的,且已经造成过实际损失:一个正在跑的程序(#12981)的在飞 PR 依赖一条现已不可达的裁决;domain:services 座位贴 §4 的 16–20 项指向一条读不到的评论。⚠️ 但要诚实:没有证据表明任何裁决因此被错误执行 —— 代价是不可复核,不是已发生的错判。 |
| 项目长远合理性 | ⭐ 指向 C,并解释了为什么 B/D 只是缩小失效面。本仓的北极星是 contract-first:一条治理这个平台的裁决,其权威载体却住在一个第三方账号的内容面上,是把契约挂在一个自己不控制的钩子上。⚠️ 但 C 与本仓另一条成文纪律直接张力:裁决往往先于实现存在(第三档「带前提的裁决」就是这个形状),而 C 要求它落树才算数 ⇒ 会出现「已裁但尚不可引用」的空窗。 |
| 防 AI 写错 | ⭐⭐ 本棱最重,且它把问题重述得更准。真正的失效不是「裁决丢了」,而是丢得静默:引用文本照常存在、assignee 照常渲染、没有任何巡查变红,而以被压制账号运行的席位连读后回读都会「确认成功」。⇒ 任何答案都必须让**「这条裁决还在不在」可机械判定**;⛔ 一个只靠人记得去看的方案(D 的镜像纪律)在本棱上不及格。⚠️ 注意 #13634 的巡查(A)已经覆盖了「检出」那一半 ⇒ 本卡真正要答的是**「权威载体放哪」**,⛔ 不要用 A 的存在来论证 A 之外什么都不必做。 |
| 创业阶段不扩散需求 | ⚠️反对 C 与 D,两者都给每一条裁决加固定成本,而裁决是本仓的高频产物(总监席记录 "第 3 场:58 张全裁")。⇒ 本棱支持 A 或 B,并建议:先让 #13634 的巡查跑起来量一轮,看断链裁决的实际频次,再决定要不要为每条裁决付固定成本。 |
推荐:B,并把 C 留作解冻项。
理由:B 复用一个已经被章程要求存在的工件(总监席每场的摘要表),边际成本接近零,而它把「一条裁决是否存在」从散布在 N 张卡上收敛成一处可巡查的清单。⚠️ 但推荐必须带上它自己的弱点:该摘要表本身也是一张账号绑定的 issue ⇒ B 缩小失效面而不消除它。
C 的解冻判据(写死,免得又变成只被复述的条目):#13634 的巡查连续跑满 7 天后,若断链引用中被判定为「承载裁决」的条目 ≥ 3,即把 C 提回决策(那时有频次数据,而不是凭本次事故的印象)。
⛔ D 不推荐:两份手工维护的拷贝会漂,而本仓已经反复实测过这一点(最近一例:#13468,一个 docblock 的四个派生数字与同文件实时 self-test 全部不符,而门全绿)。
人工地板
修改「什么算已记录的裁决」是治理协议本身的变更,且落点在 governed 面(.claude/** / AGENTS.md / skills)⇒ 人工合并即人工审核,⛔ 座位不得自裁。四棱是否同向不改变这一点。
Refs
#13634(母卡:事故测量 + 巡查半边,已 pm:queue / domain:skills / p1)· #12708(靠作者身份偶然幸存的裁决台账)· #6021(§4 指向不可达评论的座位贴)· #12981(依赖不可达裁决的在飞程序)· PR #13592(唯一一份经授权的转录)
Split out by the triage seat (#6015, R+69) from #13634, which raised it deliberately as a question rather than a proposal and named it "a skills-seat call". #13634 keeps the mechanisable half (the dangling-reference patrol) and is queued independently — A does not wait on this.
The question
Why it is being asked now — measured, not hypothetical
#13634 measured an account-level content suppression in which issues and comments authored by one account became unreachable to every other seat. The consequence that makes this a protocol question rather than an incident report:
⇒ A ruling's durability is bounded by the account that happened to file the card it sits on — a property no one chose and nothing records.
⭐ What saved that incident was an accident of authorship: the director seat post (#12708) happens to be authored by a different account, so its in-body ruling ledger survived. Had that post been filed by the suppressed account, the ledger would have gone with it.
Corroboration added by the triage seat, R+69: three further cards were measured unreachable in one session's ordinary work — #13178, #12924, #13456 — each discovered only because a seat was trying to use it. Control: dozens of other cards resolved normally in the same session on the same channel.
Options
A — Status quo. Rulings live as comments on their cards; the summary ledgers (e.g. the director seat's per-session 摘要表:卡 · 权威 · 方向) stay convenience views. Zero cost.⚠️ Accepts that ruling durability is an accident of authorship.
B — Promote an existing ledger to the durable record. The director-seat charter already requires the per-session summary table; declare that the record of a ruling, with the card comment as the elaboration. Cheap (the artifact exists), and it concentrates durability on one post —⚠️ which is itself an account-bounded artifact, so it narrows the failure rather than removing it.
C — Move rulings in-tree. A ruling becomes durable only when it lands in the repository (an ADR, a docblock, a gate, a⚠️ Highest cost per ruling and the slowest, but the only option where durability does not depend on any GitHub account.
references/row).D — Duplicate on write. Every ruling comment is mirrored into a second location under a different account/artifact at the moment it is recorded.⚠️ Two copies drift, and the mirror is only as good as the discipline that maintains it — this repo has repeatedly found that a hand-maintained second copy rots.
四棱分析
domain:services座位贴 §4 的 16–20 项指向一条读不到的评论。assignee照常渲染、没有任何巡查变红,而以被压制账号运行的席位连读后回读都会「确认成功」。⇒ 任何答案都必须让**「这条裁决还在不在」可机械判定**;⛔ 一个只靠人记得去看的方案(D 的镜像纪律)在本棱上不及格。推荐:B,并把 C 留作解冻项。
理由:B 复用一个已经被章程要求存在的工件(总监席每场的摘要表),边际成本接近零,而它把「一条裁决是否存在」从散布在 N 张卡上收敛成一处可巡查的清单。⚠️ 但推荐必须带上它自己的弱点:该摘要表本身也是一张账号绑定的 issue ⇒ B 缩小失效面而不消除它。
C 的解冻判据(写死,免得又变成只被复述的条目):#13634 的巡查连续跑满 7 天后,若断链引用中被判定为「承载裁决」的条目 ≥ 3,即把 C 提回决策(那时有频次数据,而不是凭本次事故的印象)。
⛔ D 不推荐:两份手工维护的拷贝会漂,而本仓已经反复实测过这一点(最近一例:#13468,一个 docblock 的四个派生数字与同文件实时 self-test 全部不符,而门全绿)。
人工地板
修改「什么算已记录的裁决」是治理协议本身的变更,且落点在 governed 面(
.claude/**/AGENTS.md/ skills)⇒ 人工合并即人工审核,⛔ 座位不得自裁。四棱是否同向不改变这一点。Refs
#13634(母卡:事故测量 + 巡查半边,已
pm:queue/domain:skills/ p1)· #12708(靠作者身份偶然幸存的裁决台账)· #6021(§4 指向不可达评论的座位贴)· #12981(依赖不可达裁决的在飞程序)· PR #13592(唯一一份经授权的转录)