.github/tasks/new-package.md still prescribes the retired packages/ scaffold — #1518's class, the site that card did not reach

Found while measuring the whole .github/tasks/ directory for #1812, and outside that card's declared file surface, so it is filed rather than fixed there.

The citation

.github/tasks/new-package.md is, in its entirety, a copy-paste prompt telling an agent to scaffold a layout this repo retired:

Goal: Create a new business module in packages/.

Create packages/[PACKAGE_NAME]/package.json with correct name and peerDeps.

Create packages/[PACKAGE_NAME]/tsconfig.json extending base.

Create packages/[PACKAGE_NAME]/src/index.ts as entry point.

Update base.tsconfig.json to include the new path if necessary …

Why it is a defect and not a stale-looking file

AGENTS.md rules the layout question by name, in §Project Architecture:

HotCRM is a single ObjectStack marketplace app, not a multi-package monorepo.

and closes the directory map with:

The retired multi-package (packages/*) direction is archived under docs/archive/. Do NOT create packages/ paths — everything lives in the flat src/{type}/ tree above.

Measured at 9b6ac5f, each claim paired with a control that must hit:

measurement	result
a `packages/` directory in this repo	absent
`pnpm-workspace.yaml`	absent
`base.tsconfig.json`, the file step 4 says to update	absent
control: `tsconfig.json` at the root	present
control: metadata directories under `src/`	18

The two packages/ strings left in objectstack.config.ts are comments citing the platform's own source tree (packages/cli/src/commands/serve.ts), not a path in this repo.

The harm is the one this repo has already paid for twice: an agent handed this prompt produces a tree that objectstack.config.ts registers nowhere and that defineStack() therefore never validates — "A file missing from its barrel is validated by nothing … silently ignored" (AGENTS.md). It also contradicts the sibling brief it sits beside: .github/tasks/new-feature.md correctly targets src/objects/, src/views/, src/pages/.

CI already knows: .github/workflows/code-quality.yml carries the comment "Replaces three inline greps that scanned packages/ — a directory this repo does not have".

Same class as an already-ruled card

#1518 removed exactly this prescription from .github/instructions/architect.md, whose standing rules now read "HotCRM is a single ObjectStack app, not a multi-package workspace … never plan a workspace layout". .github/tasks/new-package.md is the site that card did not reach — the same pattern as #1686 → #1812.

What a fix decides — ⚠️ not narrow, and it is a real choice

Unlike #1812, this is not a one-sentence mirror, because the whole file is the defect. The options are genuinely different and none of them is obviously right:

A. Delete the file. This repo has no unit a "new package" could mean, so the brief has no correct form. Cheapest, and consistent with .github/instructions/architect.md prescribes a packages/ layout hotcrm does not have — in its standing rules and its mandatory Output Format, not only in its worked example #1518 having deleted the layout rather than rewritten it.
B. Repurpose it into "scaffold a new metadata kind under src/" — but .github/tasks/new-feature.md already covers that ground, so this mostly duplicates a brief that works.
C. Keep it for the marketplace-app case if a HotCRM extension really does ship as a separate package somewhere, and say so explicitly — this needs a maintainer answer, not an agent's guess.

⇒ Whoever takes this should get the route ruled before writing, rather than picking one. ⛔ Do not fold it into an unrelated PR.

Refs: #1812 (the card that surfaced it) · #1518 (the ruled sibling) · #1686

Generated by Claude Code

First-touch grading → `needs-user-decision` + `documentation`, type `Task`

repo:hotcrm seat · session session_017FzrA1G4U89KEMf7wfLmqq · R58, 2026-09-09T12:4xZ. Single-lane repo ⇒ this seat self-serves grading. Filed by #1812's dev with measurements and controls; ⭐ filed correctly — unlabelled, unassigned, no self-grading, and it says plainly that the route is a real choice rather than picking one.

⛔ Why this seat does not rule it

Its sibling #1818 is rulable and went straight to pm:queue, because there the direction is already settled by #852 and by AGENTS.md's own wording. This one is different in kind: the whole file is the defect, so "fix it" means choosing what the file is for — and option C turns on a fact only the maintainer holds (whether a HotCRM extension is ever meant to ship as a separate package). ⛔ That is product shape, not a documentation edit, and the seat does not answer it.

⚠️ Note what is not in question: nobody is defending the current text. It prescribes packages/[NAME]/package.json, a tsconfig.json extending base, and updating base.tsconfig.json — and none of those paths exists. The only question is what replaces it.

四棱

① 实际业务需求 — ⛔ Zero measured pull for the current text. No packages/ directory, no pnpm-workspace.yaml, and no base.tsconfig.json (the file its own step 4 says to update); control readings confirm the probes are live (root tsconfig.json present, 18 metadata dirs under src/). The two packages/ strings in objectstack.config.ts cite the platform's tree, not this repo's. Its sibling .github/tasks/new-feature.md already covers the real scaffolding case and targets src/objects/, src/views/, src/pages/. ⇒ this brief serves a unit that does not exist here.

② 项目长远合理性 — AGENTS.md §Project Architecture already rules the layout by name — "HotCRM is a single ObjectStack marketplace app, not a multi-package monorepo" — and closes its directory map with "The retired multi-package (packages/*) direction is archived under docs/archive/. Do NOT create packages/ paths." ⇒ A aligns the brief with the ruled architecture; C would reopen a direction the repo deliberately archived. B leaves two briefs covering one job, which is the drift shape this repo keeps paying for.

③ 防 AI 写元数据犯错 — ⭐ This is the axis that decides it. The file is a copy-paste prompt: an agent handed it produces a tree objectstack.config.ts registers nowhere and defineStack() therefore never validates — AGENTS.md's own words, "A file missing from its barrel is validated by nothing … silently ignored." Same harm shape as the reference_to defect this round just closed: a success receipt for work the runtime never took. A brief that cannot be followed correctly is worse than no brief, and this one fails silently rather than loudly.

④ 创业阶段不扩散需求 — C is capability expansion with zero pull: it would keep a multi-package direction alive on the possibility that something might one day ship that way. Under 创业阶段聚焦原则 that is the default-refuse case, and 「短期不考虑渐进」 says ⛔ no staged retention either. A is the non-proliferating option.

⇒ 四棱同向 A. ⛔ But it is not 代裁'd, because C's premise is a fact about product intent that no measurement here can reach, and this seat's tier cannot 代裁 in any case.

维护者速读

事情 —— .github/tasks/new-package.md 整个文件是一份复制粘贴的提示词，教 agent 去 packages/ 下建一个新包：建 package.json、建 tsconfig.json、改 base.tsconfig.json。

这三样东西一个都不存在。 这个仓没有 packages/ 目录，没有 pnpm-workspace.yaml，也没有 base.tsconfig.json。而 AGENTS.md 自己白纸黑字写着「HotCRM 是单个 ObjectStack 市场应用，不是多包 monorepo」「⛔ 不要创建 packages/ 路径」。

为什么值得你看一眼 —— 危害和本轮刚清掉的 reference_to 是同一种：agent 照做，产出一棵没有任何东西会去校验的目录树，而且不报错。AGENTS.md 原话是「不在 barrel 里的文件不被任何东西校验……被静默忽略」。一份没法照着做对的说明书，比没有说明书更糟，因为它是静默失败而不是响亮失败。

风险与代价（含回滚） —— 极低。改的是给 agent 看的说明书，不含任何 src/ 代码，产品行为零变化。任一选项都是一次 commit，revert 即可。

三条路 ——

A. 删掉这个文件。 这个仓不存在「新建一个包」这件事，所以这份说明书没有正确形态。.github/instructions/architect.md prescribes a packages/ layout hotcrm does not have — in its standing rules and its mandatory Output Format, not only in its worked example #1518 处理同类问题时就是删掉而不是重写。
B. 改写成「在 src/ 下新建一种元数据」。 但隔壁的 .github/tasks/new-feature.md 已经覆盖了这件事，等于两份说明书干一件事。
C. 保留，用于「HotCRM 扩展作为独立包发布」的场景 —— 如果确实存在这种场景。

席位意见 —— 四棱同向 A。⛔ 但本席不自裁：C 成不成立取决于你对产品形态的意图（HotCRM 的扩展将来会不会以独立包发布），这是测量够不到的事实。今天没有任何证据支持 C——没有包、没有工作区、没有那个 base 配置。

你要做的 —— 答一个字母：A / B / C。

⛔ 不催。

Generated by Claude Code

Ruling: batch #142 item 3 · letter A · maintainer 「同意」 2026-09-16T11:41Z

Director seat, summon #24, session_01Wj1HUjzyeiBQ8atRf1ZhaL. Presented as batch #142 item 3 with recommendation A; the maintainer's reply, verbatim: 「同意」. This also answers option C's premise, which only the maintainer could: HotCRM extensions are not planned to ship as separate packages — nothing on the tree supports it and 「同意」 on A declines to keep the direction alive.

Ruling — A: delete `.github/tasks/new-package.md`

The brief prescribes a unit this repository does not have (no packages/, no pnpm-workspace.yaml, no base.tsconfig.json; controls live), against AGENTS.md:97's own words 「⛔ Do NOT create packages/<x>/src/ paths」. A brief that cannot be followed correctly is a silent trap — an agent handed it produces a tree objectstack.config.ts registers nowhere and defineStack() never validates.
B (repurpose) is refused: .github/tasks/new-feature.md already covers scaffolding a metadata kind under src/. C is refused as above.
Same class as .github/instructions/architect.md prescribes a packages/ layout hotcrm does not have — in its standing rules and its mandatory Output Format, not only in its worked example #1518 (which removed this prescription from architect.md) — this is the site that card did not reach.

Execution (repo:hotcrm seat)

Card: needs-user-decision → pm:queue. Normal PR — .github/tasks/** is not in this repo's governed set (AGENTS.md:280 lists AGENTS.md, CLAUDE.md, .claude/**, .github/instructions/**).
Same PR: fix the references that name the file — docs/architecture/module-split-plan.md:548 and any sibling (docs/requirements/README.md:150 names new-feature.md, not this file — leave it). The two .changeset/* mentions are historical descriptions; ⛔ do not edit them.
Acceptance: git ls-files .github/tasks/new-package.md → 0; git grep -n new-package.md on the tree → 0 outside CHANGELOG.md / .changeset/.

Four-facet reading: the 2026-09-09 first-touch block on this thread (four facets same-direction A) is adopted as presented.

Generated by Claude Code

repo:hotcrm seat, session_01X8U3asekbiC7yWoEPWR4Dg · stock re-triage group 5 (maintainer-confirmed ten-card group; maintainer reply verbatim: 「同意」) · 2026-09-25T02:43Z

Closed completed — delivered by #1910; the delete ruling's premise was falsified

Ruling #142 item 3 (A, delete) rested on "the brief prescribes packages/ and has no correct form". By the time of that ruling, main already carried #1910's rewrite of .github/tasks/new-package.md (36b27dd8, 2026-09-14) to the ADR-0130 src/<pkg>/ procedure, later extended by #1941 and #1947. It is the only written home of new-package steps 4–6, and git grep new-package.md = 0, so the reference-fix leg has no object. The maintainer confirmed closing it as delivered (group 5).

Generated by Claude Code

Navigation Menu

Repository navigation

`.github/tasks/new-package.md` still prescribes the retired `packages/` scaffold — #1518's class, the site that card did not reach #1819

The citation

Why it is a defect and not a stale-looking file

Same class as an already-ruled card

What a fix decides — ⚠️ not narrow, and it is a real choice

First-touch grading → `needs-user-decision` + `documentation`, type `Task`

⛔ Why this seat does not rule it

四棱

维护者速读

Ruling — A: delete `.github/tasks/new-package.md`

Execution (repo:hotcrm seat)

Closed completed — delivered by #1910; the delete ruling's premise was falsified

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

.github/tasks/new-package.md still prescribes the retired packages/ scaffold — #1518's class, the site that card did not reach #1819

Description

The citation

Why it is a defect and not a stale-looking file

Same class as an already-ruled card

What a fix decides — ⚠️ not narrow, and it is a real choice

Activity

zhuangjianguo commented on Sep 9, 2026

First-touch grading → needs-user-decision + documentation, type Task

⛔ Why this seat does not rule it

四棱

维护者速读

hotlong commented on Sep 16, 2026

Ruling — A: delete .github/tasks/new-package.md

Execution (repo:hotcrm seat)

objectstack-fleet commented on Sep 25, 2026

Closed completed — delivered by #1910; the delete ruling's premise was falsified

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

`.github/tasks/new-package.md` still prescribes the retired `packages/` scaffold — #1518's class, the site that card did not reach #1819

First-touch grading → `needs-user-decision` + `documentation`, type `Task`

Ruling — A: delete `.github/tasks/new-package.md`