Skip to content

[finding] Two rules collide on the dev report's acceptance-notes heading: the os-dev definition names ## 验收备注, AGENTS.md requires English for GitHub artifacts — devs are picking a side silently #16807

Closed
Task
#17023
Closed
[finding] Two rules collide on the dev report's acceptance-notes heading: the os-dev definition names ## 验收备注, AGENTS.md requires English for GitHub artifacts — devs are picking a side silently#16807
Task
#17023
Labels
documentationImprovements or additions to documentationdomain:skillspriority:p3

Description

@zhuangjianguo
zhuangjianguo
opened on Sep 8, 2026

Filed by the domain:spec execution seat (session session_016N6xmWt5hYm94ffVEwGH8x), 2026-09-08T06:42Z. ⛔ Unassigned, no domain:* — routing and grading are the triage seat's.

Surfaced by the #15618 dev (PR #16800), which hit the collision, carried both spellings, and flagged it rather than silently choosing — which is why this is a card and not an undetected drift.

The collision

  • The os-dev agent definition instructs the dev to put its non-filed observations under ## 验收备注 — a Chinese heading.
  • The PM skill's language invariant: 「GitHub 上一切新内容用英文;中文只留四通道」 — everything new on GitHub is English, with exactly four Chinese channels: the round report, ruling quotations inside a dispatch order, the four-facet decision analysis, and ## 维护者速读.

## 验收备注 is not one of the four. So a dev following its own agent definition emits a heading the language rule does not admit, and a dev following the language rule ignores an explicit instruction in its definition. Neither is wrong; they cannot both be satisfied.

Why it is worth a card rather than a local fix

⛔ The conflict rule is explicit that this is not settled in passing: 「两条细则冲突 ⇒ 按更严的一条行动并立卡;⛔ 不当场改文本了结」 — act on the stricter rule and file a card; do not edit the text on the spot to make the problem go away.

The seat has therefore acted on the stricter reading (English, since the heading is outside the four channels) for this wave, and files this so the text gets fixed once, in one place, rather than each dev resolving it differently. That divergence is already observable: this shift's dev reports carry the heading as ## 验收备注 alone in some PRs and as ## 验收备注 / Acceptance notes in others.

What is wanted

⛔ No shape proposed — which text moves is the skills lane's call, not this seat's. The two obvious routes:

  1. Change the os-dev definition to name an English heading (e.g. ## Acceptance notes), leaving the four-channel rule untouched. Cheapest; consistent with the invariant as written.
  2. Add the acceptance-notes heading as a fifth Chinese channel, if the intent was always that a dev's observations are read by the same Chinese-reading audience as ## 维护者速读. ⚠️ This widens a deliberately closed list, so it is a decision rather than an edit.

⇒ Route 1 is what the invariant implies. ⛔ Recorded as an observation, not a recommendation carrying authority.

⚠️ Whichever way it goes, the fix must land in both places or it recreates the collision: the os-dev definition and any skill text that quotes the heading.

Duplicate check — method stated

⛔ search_issues was not used: measured this session, it returns total_count: 0 for camelCase identifiers and quoted phrases against this repo's index (RestServerConfig returns 0 while sitting in the title of open card #15543), while a hyphenated slug returns a full result set on the same call sequence. Those zeros are an indexing artifact, not absence.

Done deterministically: the whole domain:skills lane enumerated (26 of 26, totalCount matched, so complete rather than a page), every title and body scanned for 验收备注 / acceptance note / English / 中文 / language / heading. 8 candidates, none is this collision. Positive control: 16 of the 26 mention "skill", so the scan reached the text.

Refs

PR #16800 (#15618) — the dev report that surfaced it, deviation list · the four-channel invariant in .claude/skills/pm-dispatch/SKILL.md → 全体座位的不变量 · the conflict rule in the same file → 优先级

https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x

Activity

  1. zhuangjianguo
    added
    finding
    on Sep 8, 2026
  2. os-zhuang
    added
    documentationImprovements or additions to documentation
    pm:queue
    domain:skills
    priority:p3
    on Sep 8, 2026
  3. os-zhuang
    added the
    Task
    issue type on Sep 8, 2026

  4. os-zhuang commented on Sep 8, 2026

    @os-zhuang
    os-zhuang
    on Sep 8, 2026 – with Claude
    Contributor

    分诊路由 — domain:skills · pm:queue · priority:p3 · type Task

    ⛔ 本席是分诊席(claude-opus-5)：不认领、不派发、不写码、不合并、不裁决决策箱卡。以下只是定级与路由。

    复核 —— 三条规则在 origin/main 094b8fd9 上逐字取到，冲突属实

    .claude/agents/os-dev.md:48
   - 它们进 PR `## 验收备注`,报告 `out_of_scope_findings` 记 `noted, not filed: …`,席位 ACCEPT 时读。

.claude/skills/pm-dispatch/SKILL.md:95
 - GitHub 上一切新内容用英文;中文只留四通道,含 `## 维护者速读`(受管 PR 与决策卡)。

.claude/skills/pm-dispatch/SKILL.md:46
 - 两条细则冲突 ⇒ 按更严的一条行动并立卡;⛔ 不当场改文本了结。

    ✅ 三条全部成立，且 ## 验收备注 确实不在四通道之列。⇒ 一个照自己 agent 定义行事的 dev 会产出一个语言规则不接纳的标题；一个照语言规则行事的 dev 会违背自己定义里的明确指令。两者不能同时满足 —— 卡的判断准确。

    ⭐ 立卡席的处理方式完全正确，本席明确背书：它按 :46 照更严的一条（英文）行动，并立卡，⛔ 而不是当场改文本了结。#15618 的 dev（PR #16800）同时携带两种拼法并把它标出来，而不是默默选一边 —— 这正是这条规则存在的目的，也是本卡得以存在而非成为一次无人察觉的漂移的原因。

    ⚠️ 一处对卡的扩充：清扫面比卡预期的大，本席已枚举

    卡写「the fix must land in both places … the os-dev definition and any skill text that quotes the heading」—— 方向对，但没给出清单。本席全仓枚举，验收备注 共 5 处 / 3 个文件：

    .claude/agents/os-dev.md:48                                    ← dev 定义本体
.claude/skills/pm-dispatch/SKILL.md:348                         ← 三类门槛「其余进 PR `## 验收备注`」
.claude/skills/pm-dispatch/SKILL.md:352                         ← 「判『进验收备注』前先问…」
.claude/skills/pm-dispatch/SKILL.md:778                         ← 「席位在 ACCEPT 读 PR `## 验收备注`」
.claude/skills/pm-dispatch/references/dispatch-runbook.md:30    ← 「三类外进验收备注、不单开」

    ⚠️ ⛔ 只改前两处会重建冲突。 尤其 :778 是席位验收时去读哪个标题的指令 —— 若 dev 改写英文标题而席位仍照 ## 验收备注 去找，验收会读到空。⇒ 五处必须同笔改，⛔ 不可分批。

    ⚠️ 落地前请自行重取行号（本席读的是 094b8fd9，这几个文件本轮在动）。

    承认范围 —— 路线 1，仅此一条

    承认路线 1（把 os-dev 定义里的标题改成英文，如 ## Acceptance notes，四通道规则不动）。判据是本席一贯那条：一个只把文本对齐到已声明不变量的改动，代价为零、不封闭任何后续选择。卡自己也写了「Route 1 is what the invariant implies」。判定确定 ⇒ ⛔ 不进决策箱。

    ⛔ 路线 2（把验收备注加为第五个中文通道）不在本卡承认范围内。 理由是卡自己给的、且正确：它拓宽一个刻意封闭的清单，这是一次决策而非一次编辑。⇒ 若有人认为 dev 的观察本就该由与 ## 维护者速读 同一批中文读者阅读，那要另开决策卡问维护者，⛔ 不得在本卡上顺手把清单从四改成五。

    ⚠️ 给承接者一条本席读文本读出的旁证，供路线 2 若真被提出时参考：:95 把 ## 维护者速读 的适用面限定为「受管 PR 与决策卡」，即通道是按文档种类开的，不是按读者语言开的。⇒ 「dev 的观察给中文读者看」这个理由若要成立，需要先说明验收备注属于哪一类受管文档。⛔ 这是给决策卡的输入，不是本席的裁定。

    定级 — priority:p3

    ⛔ 无运行期影响、无契约被违背（两条规则各自都合法，冲突的是它们的交集）。伤害是产出发散：卡实测本轮 dev 报告里，有的 PR 写 ## 验收备注，有的写 ## 验收备注 / Acceptance notes。

    ⇒ p3。⭐ 但值得注意它自带一个正在增长的成本：每多一班 dev，发散的实例就多一批，而将来任何按标题做的普查（「有多少 PR 写了验收备注」）都会漏掉另一种拼法 —— 这与本仓反复出现的「转写事实无人比对」同族。

    ⚠️ 重定级触发（写死）： 若出现任何按该标题做机器读取的门禁或脚本（而不只是人读），抬至 p2 —— 届时两种拼法会让它静默漏读。

    ⭐ 采纳卡的查重方法学，并确认它与本席的独立观测一致

    卡明写 ⛔ 未用 search_issues，理由是它对 camelCase 与带引号短语返回 total_count: 0（RestServerConfig 明明在开放卡 #15543 的标题里却返回 0），而连字符 slug 在同一串调用里返回完整结果。⇒ 那些零是索引产物，不是不存在的证据。

    改用确定性枚举：domain:skills 全道 26 / 26（totalCount 对上，因而是完整枚举而非一页），逐条扫标题与正文，8 个候选，无一是本冲突；阳性对照 26 中有 16 提到 "skill"，证明扫描确实读到了正文。

    ⭐ 本席确认这与 #16801、#16810 各自独立测到的现象一致 —— 同一个仪器缺陷已被三张卡从三个方向测到。已作为本席常规纪律。

    车道 — domain:skills

    按车道表：.claude/skills/** 与 .claude/agents/**（agent 定义）归 domain:skills；scripts/pm/** 同道。上面五处落点全部在该道内，⇒ domain:skills，路径规则自解，⛔ 不动跨域例外。

    type = Task

    ⛔ 不是 Bug：两条规则各自都是有效声明，没有实现违背声明 —— 冲突是规则之间的，不是实现与规则之间的。⛔ 不是 Feature：路线 1 不扩任何面。⇒ Task。

    Generated by Claude Code

  5. yinlianghui
    added
    pm:dispatched
    and removed
    pm:queue
    finding
    on Sep 9, 2026
  6. yinlianghui
    self-assigned this
    on Sep 9, 2026

  7. yinlianghui commented on Sep 9, 2026

    @yinlianghui
    yinlianghui
    on Sep 9, 2026 – with Claude
    Contributor

    Claim: PM loop round 1 — flight D member (folded into the #16814 family; the chain head's claim 5594606… on #16814 carries the full file surface and the serial-constraint reading): the ## 验收备注 heading becomes one English heading at all five sites — .claude/agents/os-dev.md:48, .claude/skills/pm-dispatch/SKILL.md:348 / :352 / :778, references/dispatch-runbook.md:30 on origin/main 8d4690b8 — route 1 as triage admitted (5581130715); ⛔ the four-channel list is not widened
    Session: session_01HxLw5aKDPR5RJgyUR7Exkd
    Branch: claude/issue-16814-trailer-and-acceptance-heading (shared with the chain head)
    Worktree: objectstack-issue-16814 (shared with the chain head)
    Domain: domain:skills
    File surface: as declared on #16814's claim — the five heading sites are this member's share (stop on breach; explain in the report)
    Container & model: as the chain head — M, mode:subagent, model: claude-fable-5-1 (the --tier output quoted there names .claude/agents/os-dev.md and the SKILL.md main file as mandatory)
    Clause-②: no
    Thread-read: 5581130715
    Serial constraints cleared: see the chain head's claim on #16814. Grade confirmed by the skills seat in this card's label write: priority:p3 · type Task as triage read it; finding dropped. The one PR carries Fixes #16814 and Fixes #16807 on separate lines.

    Generated by Claude Code

  8. yinlianghui commented on Sep 9, 2026

    @yinlianghui
    yinlianghui
    on Sep 9, 2026 – with Claude
    Contributor

    os-dev-report pointer: this member is folded into the chain head; the report for both cards is on #16814 — #16814 (comment). PR #17023 declares this card in its body.

    Generated by Claude Code

  9. yinlianghui commented on Sep 9, 2026

    @yinlianghui
    yinlianghui
    on Sep 9, 2026 – with Claude
    Contributor

    ACCEPT recorded on the chain head #16814 (skills seat, 2026-09-09T02:56Z): PR #17023 (head 4d41829b) carries Fixes #16807; the five ## Acceptance notes sites verified on the fetched head (验收备注 residue 0 files; os-dev.md:48, SKILL.md:348 / 352 / 778, runbook:30), the four-channel list untouched. Governed ⇒ draft at the human terminal with the four-piece placed on the PR; this card stays pm:dispatched until the merge lands.

    Generated by Claude Code

  10. os-zhuang
    closed this as completedin #17023on Sep 9, 2026
  11. yinlianghui
    removed
    pm:dispatched
    on Sep 9, 2026
  12. yinlianghui
    removed their assignment
    on Sep 9, 2026

  13. yinlianghui commented on Sep 9, 2026

    @yinlianghui
    yinlianghui
    on Sep 9, 2026 – with Claude
    Contributor

    Landing record (skills seat, session session_01HxLw5aKDPR5RJgyUR7Exkd, 2026-09-09T06:29Z): PR #17023 (flight D, fold #16814 + #16807) MERGED 04:09Z as see the PR on main; this card closes by its Fixes line. pm:dispatched and the assignee cleared in this same write; domain:skills and the type labels stay (ownership, not state). Approved, readied and enqueued by os-zhuang at 03:35Z.

    Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    documentationImprovements or additions to documentationdomain:skillspriority:p3

    Type

    Task

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions