docs(skills): the walk-radius mandate picks the turbo task by rule, not by name

[yinlianghui]

Fixes #16987

⚠️ Placeholder spelling in THIS body. The file's own placeholder is the angle-bracket form; every occurrence is rendered here as PKG (PKG#test, PKG#test:repo) so the body survives the platform's byte rewriting. The file is unchanged in that spelling, and every byte count below is of the FILE's bytes, never of this rendering.

What changed

.claude/skills/spec-property-retirement/SKILL.md, the tree-scoped-pin checklist item's ⭐ 半径按包申报一次 bullet (lines 278–281). It named PKG#test unconditionally. It now states the picker and names the authority, in place and line-neutral.

Before (4 lines, 392 bytes — 89 · 90 · 113 · 100):

⭐ 半径按包申报一次:写进 scripts/cross-package-test-inputs.mjs 的 CROSS_PACKAGE_TEST_INPUTS 并配齐 turbo.json 的 PKG#test inputs,turbo 才 hash 得到它、CI 分片并集才看得见;walk 降在循环变量上、门禁点不出名字,那条 glob 还要一条 heldBy 证人。⛔ 半径未申报的 tree-scoped pin 不是完成的退役。

After (4 lines, 475 bytes — 120 · 118 · 119 · 118):

⭐ 半径按包申报一次:CROSS_PACKAGE_TEST_INPUTS 写进 scripts/cross-package-test-inputs.mjs,并配齐 turbo.json:声明 test:repo 脚本的包配 PKG#test:repo 且 PKG#test 不带根输入,否则配 PKG#test;分歧以 scripts/check-cross-package-test-inputs.mjs 为准。walk 降在循环变量上、门禁点不出名字的 glob 另配 heldBy 证人。⛔ 半径未申报的 tree-scoped pin 不是完成的退役。

Kept intact: the CROSS_PACKAGE_TEST_INPUTS anchor, the heldBy witness clause, and the closing completion condition ⛔ 半径未申报的 tree-scoped pin 不是完成的退役。

⛔ Not touched: scripts/check-cross-package-test-inputs.mjs and turbo.json — the gate is correct and #10315 / #10694 hold on that script. This is a documentation defect, and the fix is only in the documentation.

The 83 added bytes are paid for by deleted content, not by a re-wrap

The bullet is 4 lines before and 4 lines after; the file is 337/337 with headroom 0, and the ratchet's per-line budget is 120 bytes, so the envelope is fixed at 4 × 120. The picker plus the authority pointer do not fit inside the old wording. They are paid for by deleting the rationale clause turbo 才 hash 得到它、CI 分片并集才看得见 — the bullet's only pure-why clause, and the one clause nothing else in the file depends on (git grep for 分片 / 并集 / hash in this file returns no other bullet stating it). ⛔ No ceiling was raised, and no line was re-flowed to buy room.

Two smaller compressions in the same envelope: 那条 glob 还要一条 → 的 glob 另配, and 的 PKG#test inputs → the picker sentence itself.

Readings — all taken on this branch at c7178179 (base ee2cb6b4), not transcribed

Packages declaring a test:repo script — git grep -l '"test:repo"' -- 'packages/**/package.json' → 6:

packages/core   packages/objectql   packages/rest
packages/runtime   packages/spec     packages/types

turbo.json inputs, counted per task on this branch:

package	#test inputs	#test ROOT	#test:repo inputs	#test:repo ROOT
core	4	0	5	1
objectql	4	0	6	2
rest	4	0	9	5
runtime	4	0	8	4
spec	4	0	34	30
types	4	0	5	1

⇒ 6 of 6, zero exceptions: on every split package the root inputs sit on #test:repo and #test carries none. The rule is flat, so it is written flat — no "in most cases", no carve-out.

The authority already knows about the split, and says so itself. pnpm check:cross-package-test-inputs on this branch prints, unprompted:

OK: 28 package(s) read outside themselves, all declared, and turbo.json hashes every declared glob (6 of them on a split "test:repo" task).

And its own self-test pins the failing direction (scripts/check-cross-package-test-inputs.mjs:2569, verbatim except the placeholder substitution declared at the top of this body):

'split: a wide input LEFT on PKG#test reds even when test:repo is complete'

The playbook's blindness to the split, and its repair:

git grep -c "test:repo" -- .claude/skills/spec-property-retirement/SKILL.md
  before (ee2cb6b4): 0     control, same file: CROSS_PACKAGE_TEST_INPUTS = 1
  after  (c7178179): 1     and check-cross-package-test-inputs = 1

Gates — 16 derived, 16 run, 16 green, 0 NOT-MEASURED

Derived from the real change set, not from the dispatch list: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at c7178179, reconciled afterwards with --ran:

✓ dispatch-gates --ran: 16 derived famil(ies) accounted for — 16 run, 0 NOT-MEASURED.

Each exit code was captured after redirecting to a file, never across a pipe. All 16 exited 0. The two that speak directly to this diff:

✓ check-skill-line-ratchet: .claude/skills/spec-property-retirement/SKILL.md is 337 lines (ceiling 337; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/spec-property-retirement/SKILL.md: widest table row is 326 bytes (pin 326; headroom 0).

pnpm check:nul-bytes is green, and an independent scan of the changed file for control bytes returned no match.

Lint, narrowed and proved narrowed. ① The linted population is read from eslint.config.mjs itself: every config object's files is a ts/tsx/mts/cts/js/jsx/mjs/cjs glob, and markdown matches none — ESLint says so in its own words, File ignored because no matching configuration was supplied. ② The count is read from --format json: 1 result, 0 files actually linted, 0 errors. ③ Invariance for untouched files: this repo "never enables type-aware linting (no parserOptions.project, no typed @typescript-eslint rules) for ANY file" (eslint.config.mjs, its own comment), so a diff confined to one markdown file cannot move any untouched file's verdict. The repo-wide pnpm lint is CI's run.

No package tests are owed. The diff touches no workspace package, so there is no dependency closure to build and no affected package's test / typecheck to run. @objectstack/lint and @objectstack/formula were built only because check:doc-formula-expressions reads their dist/.

skip-changeset: .claude/** is not published — nothing in any package's files[] moves.

维护者速读(草稿)

改了什么

退役 playbook(.claude/skills/spec-property-retirement/SKILL.md)里那条「半径按包申报一次」的检查项,原来无条件点名一个 turbo 任务。现在它写出选任务的规则(声明了 test:repo 脚本的包申报在 PKG#test:repo、其 PKG#test 不带根输入;否则 PKG#test),并点名权威(scripts/check-cross-package-test-inputs.mjs,两者分歧时以它为准)。原地改写,4 行进 4 行出,文件仍是 337 行。

为什么改

这句话是判定一次退役有没有做完的条件,而它把人送去撞门禁:门禁两个方向都管——半径没落在 #test:repo 上是一红,留在 #test 上又是一红。红的消息指向 turbo.json,不指向送他去那里的那句话,读者最省力的「修法」是去改门禁,而门禁是对的。卡面自陈:同一次派发里,卡、分诊评论、派发词三份独立文本全部继承了同一个错任务名。这次改的就是那个源头。

风险与代价(含回滚)

风险低:只动一段说明文字,不动任何门禁、不动 turbo.json,行为面零变化。代价是删掉了那条「turbo 才 hash 得到它、CI 分片并集才看得见」的理由句——4 行 × 120 字节的信封装不下新规则,而抬棘轮上限需要维护者裁决,所以用删内容付账,没有 re-wrap 凑行。回滚就是 revert 这一个提交,没有迁移、没有下游依赖。

席位意见

(留空,待评审席位填写。)

你要做的

这是 governed surface(.claude/**),按规则由维护者人工确认并合并。本 PR 保持 draft,未入队、未挂 auto-merge。要看的两处:一是新句子的规则写得对不对(6/6 实测在上面),二是同意用「删理由句」来付那 83 个字节。

Generated by Claude Code

---

Generated by Claude Code

[yinlianghui]

维护者速读 — PR #17016(#16987,skills 席终稿,2026-09-09T02:30Z)

改了什么 — 退役 playbook(.claude/skills/spec-property-retirement/SKILL.md)里「半径按包申报一次」那条检查项:原来无条件写「配齐 turbo.json 的 PKG#test inputs」,现在写出选任务的规则(声明了 test:repo 脚本的包配在 PKG#test:repo、其 PKG#test 不带根输入;否则配 PKG#test)并点名权威(scripts/check-cross-package-test-inputs.mjs,分歧以它为准)。4 行进 4 行出,文件仍 337 行,每行 ≤120 字节。

为什么改 — 这句话是判定一次退役有没有做完的条件,而它把执行者送去撞门禁:6 个拆分包(core / objectql / rest / runtime / spec / types)的根输入都在 #test:repo,门禁两个方向都拦,红的消息指向 turbo.json 而不指向这句话。上一次派发里,卡、分诊评论、派发词三份文本都继承了同一个错任务名,源头就是它。

风险与代价(含回滚) — 只动一段说明文字,不动门禁、不动 turbo.json,行为面零变化。代价:删掉了原句里的理由子句「turbo 才 hash 得到它、CI 分片并集才看得见」,腾出 83 字节;这个「为什么」在它现在指向的门禁脚本里仍有陈述(第 391 / 1450 / 1500 行),没有规则丢失。回滚 = revert 一个提交。

席位意见 — 接受。本席在拉下来的分支头上复量:字节 120 / 118 / 119 / 118、337 → 337 行、控制字节 0;受管面命中 .claude/** 一条;删掉的理由句在权威脚本里有家;6/6 人口读数与分诊表一致;门禁 16/16 绿,CI 02:29Z 无红(3 项在跑)。分诊留的两种写法(按规则命名 vs 只指向权威)本席选了两者都写:规则让读者当场能做,权威指针让下次再拆任务时这句话不再过期。

你要做的 — 合并;或由 os-zhuang / hotlong 批准,本席随即入队。看两处即可:① 新句子的规则对不对(6/6 实测在 PR 正文);② 是否同意用删理由句付这 83 字节。

---

Generated by Claude Code