Skip to content

[PM seat] domain:spec-tooling — 🔀 merged into #6017 #6018

Closed
Closed
[PM seat] domain:spec-tooling — 🔀 merged into #6017#6018
Assignees
os-zhuang
Labels
pm:seatPM seat registry issue - single-writer body, index = this label

Description

@claude
claude
bot
opened on Aug 6, 2026

范围

协议工具链 / 门禁 / 生成器 + 其文档。

面:scripts/**(门禁类)、packages/spec/scripts/**、packages/spec/docs/**、packages/lint/**、content/docs/**。
⛔ 不碰:content/docs/releases/**、packages/spec/src/**/*.zod.ts、strictness-ledger 的 counts 产物。
⚠️ 与 domain:devx 按「是否围着 spec 契约转」切分;运行时门的派发侧是 packages/metadata*(他席),本车道只拥有 packages/lint 里的注册声明。
⛔ .claude/skills/** 不在本车道(已拆出 domain:skills,座位 #7623)。

当前 PM

⏳ vacant —— 座位空缺,等待接任。

  • 前任会话:session_01WocN37om5bw81JDoEEMA2e(账号 os-zhuang)
  • 注销时间:2026-08-14T07:2xZ,维护者指示下班(「你可以下班了」)
  • 在任:2026-08-12T01:1xZ → 2026-08-14T07:2xZ,自 session_01KJATVrh6V2ysutYUJigh3B(os-help)接任
  • 定时器:前任的待命巡检触发器已全部删除(最后一枚 trig_01ErKRk3mTvoBJsyxVnqrBPe),⛔ 接任者需自建
  • 看护移交:无在飞物 —— 无在飞 dev、无云卡、无队列中的 PR、无挂起的 ADR-class PR。⇒ 无需点名看护者

接任台账(移交快照,2026-08-14T07:2xZ 现取)

在飞 0 · 队列 0 · blocked 0 · 决策箱 0 · 挂起 4 · 未定级 finding 0

卡 状态 具名重启条件
#8607 pm:on-hold 无门禁要求「破坏性行为的 changeset 必须带破坏性标注」—— ADR-0087 处置门只在已经带了标注时才开火(定级由分诊完成,重启条件见其 hold 评论)
#8360 pm:on-hold ①相邻 id 真把 PR 挤出队列 ②#7464 被拿起(两个 registry 一并设计)③#6957 被重新裁决
#8133 pm:on-hold spec 18 个自足 entry bundle 无共享 chunk(见其 hold 评论)
#7443 pm:on-hold 真的有第三个值被提议进 AUTHORING_SURFACES;⛔ runtimeTypes 增长不算

发版板(2026-08-14T07:2xZ):objectstack 3(#7644 / #7100 / #6275)· objectui 0 · cloud 0 —— 无一属本车道。

他席在办、与本车道热文件相关:#8075 / PR #8230(ADR-0049 退役 external-lookup / message-queue 两族)也改 build-docs.ts 的 data/system 页表;本席 #8166 已先落 ⇒ 对方须 rebase。截至注销时 #8230 未落 main。

热文件串行队(接任者必读)

文件 有序卡片 各自认领的区域 / 落地注记
scripts/check-adr-0087-registration.mjs + docs/adr/0087-*.md #8299(9cfdddd) 处置词表 3 值 → 4 值(新增 D7 runtime-interface-only);ADR 首次写下整份词表;assertInputs 双向断言:门接受的类目与 ADR 记录的必须同集合 ⇒ 单改一侧即红。新类目须用 路径#符号 记法 —— 裸名在本仓不构成身份(实测 PackagePublishResult 同名两处,一处是 z.input<typeof …> 投影)
ADR-0087 生成投影(spec-changes.json、docs/protocol-upgrade-guide.md) #8344(b3b618b) 已测定,别再重问:无驱动的服务端合并对二者永不「陈旧但干净」(排序并集 + 纯插入 ⇒ 要么取两侧=逐字节等于重生成,要么冲突);冲突只在两条在飞条目 id 相邻时。⛔ 分片它们买不回任何一次弹出(同场景 registry.ts 也冲突)。复现法在 entries/README.md
packages/spec/scripts/build-docs.ts #7303 → #7658 → #8166(e019766) integration 类目只剩 connector,8 个死名带有意移除注释删除。⚠️ 活撞面 #8230
packages/lint/src/validate-visibility-predicates.ts #7815(cdbffca)→ #8042(df72328) *.form.ts 提及 8 → 6,零残留在 emitted 串;2 处故意保留(文件感知调用方契约),有源码级 pin ⛔ 不要"顺手清干净"
packages/lint/src/authoring-rules.ts #7220 → #7503 → #7659 → #7576(全落) #7443(hold)动 AUTHORING_SURFACES 轴本身;#7891(他席)若派回本席,动 validateSecurityPosture 块注册 ⇒ 与任何 per-rule 卡不可并行
scripts/check-*.mjs(跨包一致性门族) #8002(2473cd2) check:filter-alias-parity AST 读三处声明(spec 表 / WIRE_QUERY_ALIAS_SLOTS+WIRE_DOLLAR_ALIASES / FILTER_SLOT_QUERY_PARAMS),任一处改形状即按设计打红("读不到"也判红);已在 CI 观察到真跑并通过
packages/spec/src/shared/alias-integrity.test.ts #7889(f7dceed) ⚠️ 精确 guidance 通道对 aliases 同样优先吞噬(无活撞面,未立卡)
packages/spec/src/migrations/(entry 树) #7624 → #7927 只加 entry 文件 + gen:migration-registry,⛔ 不手改 marker 区间

常设承诺 —— 派发前必查

触发文件 点名
check-adr-0087-registration.mjs / ADR-0087 文本 #8299 落地注记(双向断言;路径#符号 记法)
ADR-0087 entry 树 / 两投影 / registry.ts #8344 测量结论 + #8360(hold):⛔ 不提「分片投影」作为修法
build-docs.ts 的 data / system 页表 #8075 / PR #8230(他席在办)
validate-visibility-predicates.ts 文案 #8042 落地注记(2 处保留有 pin)
spec alias 表 / wire 表 / query-multiplicity.ts check:filter-alias-parity(#8002)
任何「文档锚点已验证」断言 #7484(lychee.toml 设 include_fragments = "none",不存在的锚点返回 [200] ✅ OK)
packages/spec/scripts/lib/zod-graph.ts 分诊关卡时留的知识债:getter 非记忆化、depth 上限 12/16/25/25、两格均 fail-closed(#6232/#6221 已关)

本席固定纪律(交给接任者)

  1. 读数反常先怀疑自己的模式/区间;用既有邻居做阳性对照;⛔ 不用宽区间归因。
  2. 计数不透过分页器(grep -c/wc -l);发布计数前找仓内同量 pin 对账。
  3. 注册表里的自述字段最不可信 —— 以运行的代码为准。
  4. 族/计数类断言在派发时刻重新枚举;PM 的更正同样是线索,dev 应独立重跑(Observation: build-docs.ts integration category lists 9 page names with no emitting module — only connector produces a page #8166 实证:卡面 9/228 均错,实为 8/230)。
  5. 「删除残留」类卡先确定谁写这个文件(git check-attr -a 带阳性对照 + 跑一遍生成器)。
  6. 绿色 sibling 只有在任务真正执行过时才是对照(build cache 下「绿」与「没跑」同色)。
  7. 云卡报告缺席走直接验收判据(PR body 即报告,CI 结论 + 探活回包即证据)。
  8. dev 说某门「红且既存」,先自己读那道门的 workflow 再采信(Console Pin Freshness 只在发版车道阻断,普通 PR 恒 success)。
  9. SKILL 新鲜度以已知 tip 作基准(<tip>..origin/main -- <path>),⛔ 不用共享检出 HEAD。
  10. 测量先行卡:测量可以推翻卡自己的默认方向(spec-changes.json and protocol-upgrade-guide.md are unsharded generated artifacts whose merge safety rests on a LOCAL-only git driver — the merge queue rebuilds server-side, where no custom driver runs #8344 实证,结论是「不修」且比修更有价值);裁定当众下,知识落进代码注释而非只留 issue。
  11. ADR-class 卡(docs/adr/**):approval 门要求 approving review 且禁止已武装 auto-merge ⇒ ⛔ 既不清也不绕,复核后明示挂起、卡片不摘标。ADR-0087's changeset disposition has no category for published runtime TS interfaces with no metadata surface — the #8277 exemption argument is correct, unverifiable, and will be re-litigated #8299 实证:① 批准到达后该门自动重跑,旧 job URL 永远停在 failure —— 判状态看新 attempt / Checks 面板,⛔ 不看直链;② 同事件可能触发一个被 cancelled 的重复 run,属正常取代非失败。
  12. CI 红先归因再动手:枚举分支实改文件 → 与失败点比对 → 同 commit 重跑区分抖动(feat(gates): name the runtime-interface-only ADR-0087 disposition and check it (#8299) #8456 实证:attempt 1 红在本分支未触及的 check-regen-pending.mjs fixture,attempt 2 全绿 51 步)。⛔ 判抖动时不顺带声称那个 fixture 健全。
  13. list_issues 永不返回 assignees —— 清单只是候选名单,认领前每张必过完整 issue_read。
  14. 解锁扫描连评论一起扫(issue_read 的 body 被实体转义,Blocked-by: 类行可能落在评论首行)。
  15. enable_pr_auto_merge 回显两向不可靠 —— 权威信号只有 timeline 队列事件与 MERGED;队列落地 ≈ 每 PR 15–30 分钟串行,⛔ 不据「还没落」提前判异常。
  16. finding 恒 = 待首次定级,定级即离标 —— 裸 finding 数即未定级数;车道座位可附证据/前提重验,⛔ 不定级不改标。
  17. 门禁并集必须绑 commit(协议 b43f451c):dev 须在最后一次提交之后跑并集并把该次的 git rev-parse --short HEAD 抄进报告与 PR 正文;复核比对它与 PR 当前 head.sha —— 对不上 ⇒ 那份「本地全绿」是关于死树的读数,双向都不入账;正文没抄 HEAD ⇒ 按无读数处理,以门禁 job 结论为准。棘轮族恰是最后一次提交会动的那族。

本班成绩(2026-08-12T01:1xZ → 2026-08-14T07:2xZ)

10 merged · 0 返工 · 0 空派发 · 云容器 0 空转

卡 / PR 落点 一句话
#7658 / #7808 0410522 build-docs 空 section 修复
#7576 / #7886 fc87586 surfaceReason 诚实化 + 可执行测量
#7630 / #7927 3140f9c ADR-0087 v17 entry 注册
#7889 / #7952 f7dceed alias 可达性门(含非空泛化 pin)
#7815 / #8041 cdbffca 发布门按 site 判层,schema-bound form 不再被要求写 record.
#8002 / #8182 2473cd2 新门 check:filter-alias-parity(读不到=红)
#8042 / #8183 df72328 按表面命名而非按文件名,零残留在 emitted 串
#8166 / #8219 e019766 build-docs integration 类目清 8 个死名(实测惰性)
#8344 / #8359 b3b618b 测量推翻卡的默认方向:分片投影零收益
#8299 / #8456 9cfdddd ADR-0087 处置词表 D7 + 谓词,ADR 首次写下整份词表

协议版本(注销时)

SKILL tip b43f451c。本班吸收链:ab19075 principles-only → 2570c75 notes-21 → 6ff179d S+M⇒subagent → 36f6071 spec 语义 fable → b6ff173 wave-2 硬化 → 720fe14 subagent 前置快进 → 2c0b2f3 收班=状态 flush+看护移交 → 2de1be5 四条排程裁定 → ccd46760 fable 额度耗尽豁免 → 87e2faa0 七条实测读数 → 82ea3f59 findings 流水线 → b43f451c 门禁并集绑 commit。

环境备忘

  • subagent 读本地检出,每批派发前 git -C /home/user/objectstack pull --ff-only origin main(⚠️ 实测:本班开始时检出落后 222 提交,前五个 subagent 加载的是旧 os-dev 定义)。注销时检出停在 main。
  • ⚠️ 快进后 stop hook 会报「N 个未推送提交」——那是 main 的历史,远端无该分支,⛔ 不推。
  • ⚠️ 一切读数用 git -C … origin/main:<path>;dev 一律专属 worktree;本容器 Bash 禁裸 sleep 链。
  • ⚠️ GitHub API 会撞共享账号限流(本班 05:00Z–06:0xZ 连续两枪无读数)—— 限流期如实报「无读数」,⛔ 不拿旧值充数。

跨车道备忘:⛔ list_triggers 里除本席外的条目属于其它座位,前任只删了自己那一枚。

Activity

  1. claude
    added
    pm:seatPM seat registry issue - single-writer body, index = this label
    on Aug 6, 2026

  2. baozhoutao commented on Aug 6, 2026

    @baozhoutao
    baozhoutao
    on Aug 6, 2026 – with Claude
    Contributor

    审计:维护者 2026-08-06 15:2xZ 指示换任 —— 座位改为 ⏳ 待接管,前任会话 session_01559M8FVm6W6vDLABL3jvdW 自即刻起不再认领、不再派发,仅按「在飞认领由原认领者跟完」保留 #5278/#5827 一条线的收尾(等维护者裁决 → 执行落地 → 注销认领)。C 包收官全账在 #5163;新 PM 接手面与必读纪律已写入正文。

    Generated by Claude Code

  3. os-zhuang commented on Aug 6, 2026

    @os-zhuang
    os-zhuang
    on Aug 6, 2026 – with Claude
    Contributor

    接管审计(2026-08-06 15:2xZ):

    Generated by Claude Code

  4. claude
    changed the title [-][PM座位] `domain:spec-tooling`(C 包,#5163)[/-] [+][PM seat] domain:spec-tooling — 🟢 active[/+] on Aug 6, 2026
  5. os-zhuang
    changed the title [-][PM seat] domain:spec-tooling — 🟢 active[/-] [+][PM seat] domain:spec-tooling — 🟡 收官中,待接管[/+] on Aug 7, 2026
  6. os-project-manager
    changed the title [-][PM seat] domain:spec-tooling — 🟡 收官中,待接管[/-] [+][PM seat] domain:spec-tooling — 🟢 os-project-manager[/+] on Aug 7, 2026
  7. os-project-manager
    self-assigned this
    on Aug 7, 2026

  8. os-project-manager commented on Aug 7, 2026

    @os-project-manager
    os-project-manager
    on Aug 7, 2026 – with Claude
    Collaborator

    Takeover audit (2026-08-07):

    Generated by Claude Code

  9. os-zhuang
    changed the title [-][PM seat] domain:spec-tooling — 🟢 os-project-manager[/-] [+][PM seat] domain:spec-tooling — 🟡 待接管(在飞已清零)[/+] on Aug 7, 2026
  10. os-project-manager
    changed the title [-][PM seat] domain:spec-tooling — 🟡 待接管(在飞已清零)[/-] [+][PM seat] domain:spec-tooling — 🟢 在任 (session_01AZgRyPVwi1jLb1mNNuUQ9o)[/+] on Aug 8, 2026

  11. os-project-manager commented on Aug 8, 2026

    @os-project-manager
    os-project-manager
    on Aug 8, 2026 – with Claude
    Collaborator

    巡逻记录 02:0xZ —— 车道纵深实测:6 张 pm:queue 里当前可派的是 0 张,原因各不相同

    本轮为「是否加派第二个 dev」做的评估,结论是不加派,但过程中的四条事实值得留在卡上,免得下一个拉车道的座位重新踩一遍:

    教训一条,写在纪律㉓ 旁边:pm:queue 不等于可派。 本轮六张卡里两张(#5873、#5757)的标签比它们的前提新,且其中一张的晋级是本座位自己做的 —— 派发前的活状态实读必须包含「晋级理由本身是否还成立」,而不只是「正文指认的行还在不在」。

    静默死亡观察维持 2/3,本轮未增。 02:0xZ 复核 PR #6491:26 项检查 25 项完成全绿,TypeScript Type Check 仍在跑 —— CI 未收敛,dev 未报告属预期行为,此时探活会污染这个观察,故不记数据点、不发探针。02:42Z 那班巡逻在收敛后再判。

    Generated by Claude Code

  12. os-project-manager commented on Aug 8, 2026

    @os-project-manager
    os-project-manager
    on Aug 8, 2026 – with Claude
    Collaborator

    02:2xZ —— 加速档:台账转 3 落地 / 3 在飞,并新增一条常设授权

    ⚑ 常设授权(维护者 02:2xZ 批复):凡本座位派出、CI 全绿且已通过 PM 验收的 PR,由本座位自行落地,不再逐次请示。

    ⚑ 落地机制的事实更正,写在授权旁边:本仓不允许直接合并 —— PUT /pulls/{n}/merge 被仓库规则拒为 405 Changes must be made through the merge queue。所以「合并」这个动作在本仓的正确形态是开启自动合并 / 进入合并队列,不是调 merge。前三单之所以在本座位未介入的情况下落地,机制就是这个队列。后续座位不必再把「谁来合」当悬案。

    在飞三单(彼此文件面两两不相交,已逐对核过)

    Issue 分支 / 面 备注
    #6420 PR #6491 已进合并队列(squash 自动合并) 验收通过,CI 26/26;落地即解锁 #6484
    #6260 .github/workflows/** Check Changeset 触发/读标签 修法方向锁死:读实时标签或改 labeled 触发;⛔ 禁止「让作者手更快」——18:31Z 证据已证伪
    #6350 v17 存量 227 条 breaking changeset 一次性回补审计 ⛔ 明令不得改成回溯全量门禁(#6129 拒绝过的形状);2 条抽样疑似必须逐条落到结论
    #6484 file-description.ts + build-docs.ts 契约(fromCategory) 条件式派出:建分支前必须确认 #6491 已在 main 上,否则从旧基线起手必造假冲突

    新增车道纪律(即刻生效):同一时刻只允许一个「全语料重生成」面在飞。#6484 占用中,故 #6374 挂起,待其落地再派。理由是 #6420/#6473 那次同页相撞的实测代价 —— 生成物只能在合并树上整体重跑,不能文本合并,并行两个重生成面等于给自己造串行税。

    接口裁决沿用而非重开:#6484 的路线 A(FileDescriptionContext 加 fromCategory)由分诊 01:54Z 定为有约束力,本座位照办并写进派单(倾向 B 者必须停下上报)。其验收判据同样照抄:判据是「没有一处仍以纯文本落地」,不是「九处全成链接」——5/9 无目标页者按 #6229 回退成代码段。

    Generated by Claude Code

  13. os-project-manager commented on Aug 8, 2026

    @os-project-manager
    os-project-manager
    on Aug 8, 2026 – with Claude
    Collaborator

    02:3xZ —— #6260 空派,前提在派单前 7.5 小时就已死。台账 3 落地 / 2 在飞 / 1 空派。

    发生了什么:#6260 于 02:11Z 派出,dev 首件事做 stale-premise 检查即证伪 —— 该缺陷早在 08-07 18:48:35Z 由 domain:devx 座位以 #6378 / PR #6429(35353bd,Check Changeset 的「结算读」)修掉,比本座位 00:47Z 晋级早约 6 小时、比 02:11Z 派单早约 7.5 小时。本座位复核确认后已关闭 #6260 为 #6378 的重复件。零 PR、零改动,成本是一次 dev 启动(约 13 万 token)。

    ⛔ 新增硬性纪律㉔ —— 落点路径的 list_commits 是派单前置,不因加速省略。

    这一步不是可选的尽职调查,它是唯一能发现此类死亡的检查:一次 list_commits --path .github/workflows/pr-automation.yml 就会把 35353bd 直接摆出来。本座位派 #6420 时做了(前提成立,派对了);派 #6260 / #6350 时在加速档里省掉了。「加速」允许并行更多 dev、允许缩短巡逻间隔,不允许削掉派单前置检查 —— 省掉的那一次 API 调用,代价是整个 dev 轮次。

    尤其反讽的是:本座位在 02:0xZ 的巡逻里刚把「pm:queue 不等于可派,标签可能比前提新」写进车道观察(#5873、#5757 两例),然后在下一轮自己犯了同一类错误 —— 区别只在于前两例是别人的晋级、这一例是自己的。规则对自己同样生效,这是第 3 例,不必再等阈值。

    ⛔ 附带更正一条已发布的错误归因(本座位 18:31Z 写在 #6260 上的):我当时断言「该步读的是事件载荷而非活标签」。实测不成立 —— 它读的是活标签,只是读得太早(~+10s,标签落在 +10..45s);我引为「标签落地 8 秒后仍判红」的 18:30:07Z 是判定步骤在消费更早那次活读的结论,不是一次新读取。症状与修法方向都对,机制错了。更正已留在 #6260 线程,防止其他座位继续引用。

    #6350 同批派出、同样漏做该检查,其 dev 已带 stale-premise 条款,若同样死亡则再记一次;不预判,等报告。

    当前在飞:#6350(存量回补审计)、#6484(fromCategory 契约,已条件式派出)。#6491 在合并队列。#6374 仍按「同一时刻只允许一个全语料重生成面在飞」挂起。

    Generated by Claude Code

  14. os-project-manager commented on Aug 8, 2026

    @os-project-manager
    os-project-manager
    on Aug 8, 2026 – with Claude
    Collaborator

    03:5xZ —— 静默死亡观察达到阈值 3/3,给出生产者侧改法。台账 5 落地 / 1 在飞 / 1 空派。

    第 3 例,且是第一例确证

    #6484 的 dev(PR #6534)在 CI 25/25 全绿(03:35:30Z 收敛)后 22 分钟未交报告。发探针,运行时回执是 had no active task; resumed from transcript —— 它已经停了,不是在做本地验证。前两例(#6419、#6452)是从沉默推断的,这一例是运行时直接给的答案。

    但把账算全:同一步上有 2 例存活

    dev 在「等 CI 收敛」这一步 结果
    #6419 死 探针复活后报告干净
    #6452 死 探针复活后报告干净
    #6420 活 自己等到 26/26 收敛、逐 job 读真实结论后交报告
    #6350 活 同上,22 项收敛后交报告,并附 CI 步骤级时间线
    #6484 死 本次

    3 死 / 5 例,60% 死亡率,不是必死。这一点很重要 —— 它决定了改法不能是「禁止 dev 看 CI」。

    生产者侧改法(提交给 os-dev 定义)

    两个存活样本恰恰证明了看 CI 是有价值的:#6420 的 dev 因此发现 ESLint job 携带着本地 pnpm test 覆盖不到的家族门禁;#6350 的 dev 因此拿到了步骤级时间线,成为它证伪 #6260 的关键证据。所以不能一刀切掉。

    改法:把「报告」与「等收敛」解耦,顺序强制。

    1. 推完 PR、开完草案 PR 立即交结构化报告 —— 此时报告已包含全部本地门禁结果与实现结论,是整轮工作的主要产出。
    2. 之后若要继续观察 CI,作为补充报告再发一次(存活时得到 docs-gen: 括号里的裸源码路径既不成链接也不成代码 —— 4 处、3 张已发布参考页(#6224 之后残留的 lookaround 对) #6420/[finding] #6148 的门禁只判 diff,v17 列车已有的 227 条 breaking changeset 从未被比对过 —— 抽样已见 2 条疑似同形漏登记 #6350 那种额外价值;死亡时损失的只是补充,不是整轮)。
    3. CI 收敛的核验归 PM —— 本座位每轮巡逻本来就独立读 check runs,dev 那次等待对 PM 不产生任何 PM 不会自己拿到的信息。

    现状下这 60% 的死亡率赌的是整轮报告;改后赌的只是一份增量补充。今天已有两次为此花掉探针往返(#6419、#6452),本次是第三次。

    顺带修正本座位一条推断

    我此前认为「无完成通知 = agent 仍在运行」,因而以此为由推迟探活。本次证明该推断不成立 —— agent 可以停在无活动任务状态而不产生通知。今后判据改为:CI 收敛 + 超过 20 分钟无报告 ⇒ 直接探针,不再拿「没收到通知」当活着的证据。这条改动同时避免了反向错误:探针本身是无损的,活着的 agent 只会回一句「还在验证」。

    Generated by Claude Code

  15. os-project-manager commented on Aug 8, 2026

    @os-project-manager
    os-project-manager
    on Aug 8, 2026 – with Claude
    Collaborator

    04:0xZ —— #6484 验收通过进队列。新增纪律㉕。台账 5 落地(+1 在队列)/ 0 在飞 / 1 空派。

    ⛔ 纪律㉕:为达成验收判据而必须做、却落在裁决契约之外的改动,实施前先申报

    一段话说明扩的是什么、为什么绕不开,然后继续做。这是告知,不是请示,不产生往返成本。

    来源是 #6484 的实例:dev 认出了派单里的 stop-and-report 条件(它当场写下过这句话),随后把 tripwire 重读成「关于已发出产物的性质」—— 而它的设计让产物零死链,于是判定未触发,并把「stop-and-report 条件未触发」写进 PR 正文。字面站得住,但不是意图。更关键的是第二半:它把改动记录得极充分(独立章节 + 一段源码注释 + 专为量它而设的回退组),却没有一处称其为扩面。记录充分 ≠ 申报。

    规则洞的形状值得记下来,因为它是结构性的:Prime Directive #10 管范围外缺陷(另立单、不修),needs_decision 管欠定的契约选择 —— 而这次是「范围内且承重」,两条都读作不适用,于是从缝里掉下去。㉕ 补的正是这条缝。

    派单措辞同步收紧(一行):「新发出的路由无对应页面」→「无对应页面的路由,无论已发出还是你这次改动会触及」。它单独拦不住这次(那句话作为「已发出产物的性质」本就准确、且正确地从未触发),但堵掉了被用上的那个字面读法。

    数据卫生:一个错数字穿过了三层

    派单里「9 处中 5 处目标页不存在」是错的,实测 4 处。链条是:issue 正文的表本身是 5 有 / 4 无 → 分诊评论把它写成「5 of the 9 targets have no published page」→ 本座位照抄进派单当硬事实 → dev 按派单要求自行复核并纠正。dev 那一环工作正常,前两环没有。推论:派单里引用的每个数字,要么自己量过,要么明确标注为「转述,待实施者复核」 —— 不要把转述写成断言。

    ⚠️ 下一轮 #6374 的陷阱(㉔ 预检已做一半)

    #6374 释放条件是 #6534 落地(同一时刻只允许一个全语料重生成面在飞)。但释放 ≠ 可派:它的落点 build-docs.ts 自 #6374 立单(08-07 15:24Z)后已被改动两次 —— #6377(15:53Z,顶层长枚举移入 Allowed Values + 联合变体计数,正是 formatType 那一带)与 #6534(本次,+83/−29)。#6374 讲的是 INLINE_KEY_LIMIT 沿数组/Record/联合无预算下钻,与 #6377 的射程高度相邻。派它之前必须按 ㉔ 重跑逐路径检查,并确认 #6377 没有顺带改掉它的前提 —— 这正是 #6260 空派的同一形态,那次代价是一次 dev 启动。

    车道新到

    #6539(由 #6484 的 dev 立)—— build-docs.ts 在 import 时执行,其路由/分页逻辑没有可转红的单测覆盖:实测删掉一处真实的正确性校验,pnpm test 仍 57/57 全绿,而产物多出 4 条死链。已由必需的 lychee Check Documentation Links 覆盖住实害且当前无死链,故 dev 自评观察级并只打了 finding + domain:spec-tooling、未打 pm:queue —— 定级留给分诊,这个分寸是对的。

    #6497 仍未定级(live bypass:无箭头的改写表可买到 not-required 豁免,今日 4 条 breaking changeset 正走这条),继续观察其分诊延迟。

    Generated by Claude Code

  16. 8 remaining items

  17. os-help
    assigned
    os-help
    and unassigned
    os-project-manager
    on Aug 10, 2026

  18. os-help commented on Aug 10, 2026

    @os-help
    os-help
    on Aug 10, 2026 – with Claude
    Collaborator

    Takeover audit (2026-08-10T01:2xZ):

    Generated by Claude Code

  19. os-help commented on Aug 11, 2026

    @os-help
    os-help
    on Aug 11, 2026 – with Claude
    Collaborator

    Round 13 opened at 05:1xZ — ledger delta, so the body's In flight: 0 is superseded until the next body write.

    ㊴ confirmed again — this is the eighth refill

    #7503 was not in the lane at my 05:0xZ query and was at 05:1xZ; triage attached domain:spec-tooling at 05:07:49Z. The card itself was filed 04:09Z. ⇒ The gap between "queue drained" and a live card was under ten minutes, and the drained reading was correct when taken. ⛔ Never report a queue from cache; re-query at the moment of writing.

    ⚠️ Tool trap worth recording: list_issues(labels: [...]) is OR, not AND. ["domain:spec-tooling","pm:queue"] returns every pm:queue card in the repo — 28 rows across every lane, which reads like a lane query and is not one. Query the domain label alone and read each row's pm:*.

    ⚠️ Deviation from triage's suggestion, disclosed (㉕)

    Triage suggested S, sonnet. Dispatched opus, mode:cloud. Not a re-grade — importance is triage's single channel and is unchanged — but sizing/model is the seat's, and two premises failed re-measurement at dispatch time:

    1. ㊽ — "family precedent is warn" does not survive re-enumeration. validate-security-posture.ts has 12 rule ids / 14 emit sites: 11 error, 2 warning, 1 info. The file's own criterion (authoring-rules.ts:1055-1058, ADR-0090 D7) is "every error rule mirrors a runtime enforcement point … per ADR-0049 this is not advisory security" — and this rule mirrors assertControlledByParentWrite answers a metadata defect and a missing row with the same 403 PERMISSION_DENIED "requires edit access to its master record" #7474's 422 exactly. The three non-error rules are the three that mirror no hard runtime refusal. warn is the adjacent rule's precedent, not the family's.
    2. ㊼ — the block's surfaceReason is a coverage claim that assertControlledByParentWrite answers a metadata defect and a missing row with the same 403 PERMISSION_DENIED "requires edit access to its master record" #7474 is evidence against. validateSecurityPosture registers once as a block (authoring-rules.ts:1059-1073) with surfaces: CLI_ONLY, so a new rule silently inherits it. The stated reason is that plugin-security's ADR-0094 registerAuthoringGate already enforces "the same OWD posture rules" at runtime-publish — but if it caught this shape, assertControlledByParentWrite answers a metadata defect and a missing row with the same 403 PERMISSION_DENIED "requires edit access to its master record" #7474 would not have needed a runtime write refusal. Under CLI_ONLY the new rule would not fire on the runtime-publish path, which is the exact path the card names as mattering most ("AI-authored metadata" — agents publish, they do not run os lint).

    That second item is a fact about the registry that a mechanical pass would inherit rather than measure, which is what makes this M rather than S. Same axis this lane moved six rule ids across in #7220 hours ago.

    ⛔ The dev is explicitly forbidden from flipping the block to CLI_AND_RUNTIME (12 rules' surface in one PR; authoring-rule-wiring.test.ts:363 polices the field) and from restarting #7443 — whose hold condition remains a third value actually proposed for AUTHORING_SURFACES. If the gate gap is real it comes back as a separate finding, unassigned, for triage to grade.

    ⚠️ Noted, not acted on: #7443's body still carries Blocked-by: #7220, and #7220 landed as PR #7479. The hold comment supersedes it; leaving the filer's text rather than churning a held card, but the next reader should not be misled by that line.

    Also this round

    Generated by Claude Code

  20. os-help commented on Aug 11, 2026

    @os-help
    os-help
    on Aug 11, 2026 – with Claude
    Collaborator

    Round 13 review — report ACCEPTED, PR #7574 undrafted and armed. And the brief I wrote contained a wrong count, which the dev caught.

    ⚠️ First, the seat's own error — new discipline 51

    My dispatch brief stated the security-posture family as 14 emit sites / 11 error. The dev re-measured and reported 15 / 12. Re-verified independently just now on origin/main:

    error    12
warning   2
info      1
total    15

    The dev is right and I was wrong. Mechanism: I derived the count from a compound grep … | head -80. The head truncated the output before I counted it, and I then published the truncated number as a measurement. Nothing in the output says it was cut — that is the whole problem.

    Worse, the repo already held a positive control that would have caught it instantly: validate-security-posture.test.ts:856 pins expect(pushedRuleIds()).toHaveLength(15). The file counts itself, and I did not check my number against the file's own pin. The dev did, which is how the discrepancy surfaced.

    51 — never derive a COUNT through a pager. grep -c, not grep | head. A truncating filter in a counting pipeline produces a number that is wrong in one direction only (low) and carries no signal that it was truncated. And before publishing any count into a brief, look for an in-repo pin of the same quantity — a length assertion, a snapshot, a ratchet — and reconcile against it. This is ㊾'s sibling: ㊾ says establish which quantity the line prints; 51 says establish that you saw all of it.

    ⚠️ This is the second time this shift a count I supplied survived for the wrong reason (#7465's 38 was the first). Both times the dev found the right ground independently. That is the system working — but the brief is supposed to reduce the dev's work, not add a falsification task to it.

    Review verdict: ACCEPT

    Both mandated measurements were answered explicitly, neither deflected:

    ① Severity error — argued from the file's own ADR-0090 D7 criterion, not from head-count, and strengthened with something I had not supplied: the runtime refusal is both halves, not one — writes get 422 (#7474) and computeControlledByParentFilter returns RLS_DENY_FILTER, so the object is unusable, not merely locked down. It also distinguished the adjacent warning correctly: security-master-detail-ungranted is advisory because of per-permission-set nuance it cannot adjudicate; this defect is self-contained in the object document. Independent corroboration it did not go looking for: content/docs/permissions/authorization.mdx:328 states the same criterion in a file that does not cite authoring-rules.ts.

    ② registerAuthoringGate → branch 3, measured. The gate lives in object-posture-gate.ts (127 lines, read in full): it implements exactly R1 env-tighten-only and R2 external ≤ internal, reads only sharingModel and externalSharingModel, and never reads fields — so it structurally cannot see a relation. controlled_by_parent is additionally excluded from its OWD_WIDTH by design. ⇒ block stays CLI_ONLY; surfaces, AUTHORING_SURFACES, CLI_AND_RUNTIME, runtimeTypes and plugin-security all untouched; #7443 not restarted. Gap filed as #7576 with the full 13-rule map (the gate covers 1 — security-external-wider-than-internal, which is R2), and deliberately narrow about what it does not claim.

    Accept bar met exactly as specified: one positive control that fires (plus a second), and three separate negative controls, one per fallback step — not one combined. The dev also stated why three: "a rule mirroring only step 1 passes a single combined control and fails these two."

    Three things it found that the card, the triage note and my brief all missed:

    • resolveCbpRelation's pick requires pred(f) && ref(f) — every step also needs a reference target. Neither the issue body nor my brief says this; it came from reading the runtime instead of the prose.
    • The runtime reads f?.required for truthiness, not === true. Mirroring a gate means mirroring its coercions.
    • ⚠️ The security linter's own clean-stack fixture was an instance of the defect — leave_item declared controlled_by_parent with no fields at all and asserted toEqual([]). Fixed in the same commit. A negative fixture that is itself defective is the one failure a new rule's test suite cannot detect from inside.

    One deliberate divergence, documented at the helper rather than taken silently: the runtime accepts reference ?? reference_to ?? referenceTo; the lint reads only reference, because the aliases do not parse under the strict schema and copying the fallback would restore exactly the inert branch #5017 removed from this same file. Correct call, and correctly disclosed.

    Disclosed scope addition (㉕), accepted: content/docs/permissions/authorization.mdx — one clause, because that page enumerates this linter's error rules by hand and the docs-drift check flagged it. In this lane's surface, and ⛔ content/docs/releases/** untouched.

    Landing

    CI converged before the flip, verified by check runs not by webhook (㉞): 26 success / 1 skipped (Console Pin Gate), mergeable_state: clean. File list read before undrafting: 5 files — changeset, authorization.mdx, index.ts, the rule, its tests. ⛔ No docs/adr/**, no content/docs/releases/**. Real changeset (@objectstack/lint minor) — correctly not skip-changeset, since a new authoring rule is a public behaviour change (#7479's lesson).

    Undrafted → enable_pr_auto_merge (SQUASH). ⚠️ Empty-field receipt for the fifth consecutive time; queue branch read empty at 06:24Z, which on #7479 also immediately preceded the enqueue notice. ⛔ No reflex disable→enable. Subscribed; landing check armed 07:06Z.

    Ledger

    17 merged · in flight 1 (#7503 / PR #7574, armed) · queue 0 · needs-user-decision 2 · on-hold 1 · rework 0 · empty dispatch 0.

    Findings produced this lane now also: #7576 (⛔ triage's to grade — not recorded here as anything but filed).

    ⚠️ Environment note from the dev, worth carrying: it could not create its worktree on the designated branch name because the shared primary checkout already had claude/issue-7503-lint-cbp-without-relation checked out, and switching it off was blocked by the permission classifier. It worked in a dedicated worktree on a local branch and pushed with an explicit refspec. No git stash anywhere, nothing written in the shared checkout. ⇒ The shared checkout can be left sitting on a previous dispatch's branch, which the next dev inherits as an obstacle. Not a defect in this PR; a container-state fact the next dispatch should expect.

    Generated by Claude Code

  21. os-help commented on Aug 11, 2026

    @os-help
    os-help
    on Aug 11, 2026 – with Claude
    Collaborator

    Audit — seat post recompressed to the six-section template (#7583), in place, same incumbent. Not a takeover: session session_01KJATVrh6V2ysutYUJigh3B continues.

    Archive pointer: the previous body is the immediately preceding body revision of this issue, last written 2026-08-11 ~06:4xZ. Nothing was dropped silently — the mapping is below.

    Why now

    efc3c0e0d (PR #7698, ruled batch) landed the six-section body template — 范围 | 当前 PM | 继承台账 | 热文件串行队 | 常设承诺 | 说明, every section a current value, no section growing over time — and the paired rule that shift narrative does not belong in the body (#7583). This post had become exactly the banned shape: a per-card ledger table that only grew, a discipline list at 51 entries, and an operational-facts section accreted event by event. #6019 hit ~61 KB and broke single-read tooling; check-half-states.mjs H6 now flags any pm:seat body over 10 KB.

    ⚠️ I did not notice this from a webhook or a nag — it came out of the standing staleness check. The protocol had moved one commit ahead of the copy I was operating on, and that commit changed how this very post must be written.

    Where the old content went

    old section now
    per-card ledger table (22 rows, append-only) 继承台账, as a current value — 在飞 0 / 队列 0 / 决策箱 1 / 挂起 1 / 未定级 finding 7. Per-card status lives on the cards and PRs, which is where the protocol says it lives; round totals go to the step-9 round report
    「Protocol refresh」段 deleted — it duplicated SKILL.md, which is exactly the copy that went stale under me
    disciplines ㉔–51 the general ones were absorbed into the protocol by #7698 itself (my 9/9 empty-receipt readings are now notes 21; the whole-lane single-query inventory shape is now codified as #7518 lesson 5). The lane-specific five survive verbatim under 常设承诺
    「Operational facts」 split: environment-shaped ones → 说明; the four that must fire at dispatch time → 常设承诺 → 派发前必查, which is the form #7518 lesson 3 requires (a check hanging on the dispatch action has a reader; a note in prose does not)
    findings-produced list dropped as a list. Grading is triage's channel and the label query is the index; the four that must be named in a dispatch order are now in 派发前必查 with their trigger files

    Two things this surfaced that are now on the post

    A hot-file serialization queue this lane actually needs. Four files carry multiple cards, and one pair is genuinely unparallelizable: #7576 would change validateSecurityPosture's block-level surfaces — 13 rules in one field — so it cannot run beside any per-rule card in authoring-rules.ts. That was true before today and lived nowhere; it now has a named section, per #6644's measurement that the region column is what makes parallel dispatch safe.

    ⛔ .claude/skills/** left this lane. #7548's second ruling split domain:skills out of domain:devx (seat #7623). My old body carried "anything touching .claude/skills/pm-dispatch/** dispatches at claude-fable-5" as a lane rule — that surface is not this seat's at all now; such cards go to #7623 and to the maintainer's 专题 channel as ADR-class. Removed rather than left as a stale claim on someone else's territory.

    Corrections to my own practice, from the same batch

    Generated by Claude Code

  22. os-zhuang commented on Aug 12, 2026

    @os-zhuang
    os-zhuang
    on Aug 12, 2026 – with Claude
    Contributor

    Takeover audit (2026-08-12T01:08Z):

    Generated by Claude Code

  23. os-zhuang
    changed the title [-][PM seat] domain:spec-tooling — 🟢 os-help[/-] [+][PM seat] domain:spec-tooling — 🟢 os-zhuang[/+] on Aug 12, 2026
  24. os-zhuang
    assigned
    os-zhuang
    and unassigned
    os-help
    on Aug 12, 2026
  25. os-zhuang
    changed the title [-][PM seat] domain:spec-tooling — 🟢 os-zhuang[/-] [+][PM seat] domain:spec-tooling — ⏳ vacant[/+] on Aug 14, 2026

  26. os-zhuang commented on Aug 14, 2026

    @os-zhuang
    os-zhuang
    on Aug 14, 2026 – with Claude
    Contributor

    Shift-end audit — seat vacated, session_01WocN37om5bw81JDoEEMA2e signing off at 2026-08-14T07:2xZ on the maintainer's instruction (「你可以下班了」).

    Exit sequence, per the seat-post protocol:

    1. New dispatch stopped — no card was claimed after ADR-0087's changeset disposition has no category for published runtime TS interfaces with no metadata surface — the #8277 exemption argument is correct, unverifiable, and will be re-litigated #8299.
    2. Per-item disposition — nothing to hand over: in-flight 0, no dev subagent, no cloud card, no PR in the queue, no ADR-class PR hanging. The hot-handover clause (2c0b2f3) therefore names no successor watcher, because there is no in-flight item to watch.
    3. Lane enumerated into the handover ledger above (fresh read at sign-off, not cached): queue 0, decision box 0, pm:blocked 0, on-hold 4 (No gate requires a breaking-behavior changeset to carry the breaking annotation at all — the ADR-0087 disposition gate only fires once one already does #8607 / migrations/registry.ts still text-merges: two ADR-0087 entries with adjacent ids conflict server-side, which is the residue #7297's source sharding could not reach #8360 / @objectstack/spec ships 18 self-contained entry bundles with no shared chunks, so a consumer mixing entry points gets different ConnectorSchema / defineConnector objects (measured) #8133 / [P3] Generalise the lint rule surfaces axis to an open N-surface dimension (#4463 P1 introduced a closed two-value one) #7443, each with its named restart condition), ungraded finding 0. Release board 3 / 0 / 0, none in this lane.
    4. Seat post rewritten to ⏳ vacant with the full ledger — predecessor session id, sign-off time, queue snapshot, hot-file serialization queue (the part a successor most needs before dispatching anything), the pre-dispatch cross-check table, 17 lane disciplines, and the cross-lane note.
    5. Timers cancelled: the standby patrol trigger trig_01ErKRk3mTvoBJsyxVnqrBPe was deleted. ⛔ Every other entry in list_triggers belongs to another seat and was left untouched. A successor arms its own.

    Shift result: 10 merged · 0 rework · 0 empty dispatch · 0 idle cloud container. Cards and landing commits are tabled in the seat post.

    Three things a successor should know that are not obvious from the card list:

    Generated by Claude Code

  27. hotlong
    changed the title [-][PM seat] domain:spec-tooling — ⏳ vacant[/-] [+][PM seat] domain:spec-tooling — 🔀 merged into #6017[/+] on Aug 14, 2026

  28. hotlong commented on Aug 14, 2026

    @hotlong
    hotlong
    on Aug 14, 2026 – with Claude
    Contributor

    Lane merged (maintainer ruling 2026-08-14, verbatim, untranslated: 「spec 任务现在少了，是不是 spec 相关的车道也可以合并了。」→「同意」on the merge proposal). domain:spec-tooling is now a routing sub-label of the spec seat (#6017) — this seat post takes no further takeovers. Lane state at merge: four open cards, all pm:on-hold (#8607, #8360, #8133, #7443), zero queued, zero in flight; they stay in place and their restart conditions now wake into #6017. The spec exclusivity ruling (「其他人不得接手」) covers this label from now on. Do not re-seat here; claim work via #6017.

    Generated by Claude Code

  29. os-steve
    closed this as completedon Aug 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

pm:seatPM seat registry issue - single-writer body, index = this label

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions