docs: CLAUDE.md becomes a pointer to AGENTS.md — the inlined never-miss digest leaves

[os-justin]

Fixes #17162

Program anchor: #17161 (that card stays open — this repo is one of several).

Maintainer direct instruction, 2026-09-09, verbatim and untranslated:

「首先要明确 hotcrm 是元数据应用，平台功能应该在 objectstack中开发。 其次 claude.md 是不是直接让他阅读 agents.md 即可，没必要维护两套。所有仓库都有类似的问题」

What landed

CLAUDE.md goes 41 lines to 5: a title and one bold sentence naming AGENTS.md as the single
source of truth to read before the first edit. The four inlined never-miss rules leave the file.
.claude/hooks/** is untouched — the hooks keep enforcing worktree-first, the stash ban and the
process-kill ban exactly as before.

Each of the four was verified to survive in AGENTS.md by its own sentence, not by a word
hit, before deletion:

rule leaving CLAUDE.md	anchoring sentence in AGENTS.md
claim-first	AGENTS.md:393 — "Claim the issue BEFORE you write any code." Every agent here shares one GitHub identity, so the assignee field is only a presence bit; the identity record is the Claim: comment… (runs to :412, and carries the dispatch/seat split and the ⛔ never-reassign clause the digest compressed)
worktree-first	AGENTS.md:217 — Prime Directive 11. "Worktree-first — never edit on the shared main checkout." …Before your first file edit, be in a dedicated worktree on a feature branch… (names both guard hooks and OS_ALLOW_MAIN_EDITS=1)
the git stash ban	AGENTS.md:318 — "⛔ git stash is the sharpest thing the worktree does NOT isolate — never run a bare git stash push/pop." (names guard-shared-stash.sh, OS_ALLOW_STASH=1, and the collision-free replacements)
the kill-by-name ban	AGENTS.md:528 — "⛔ One process table per container: kill only a PID you recorded, never a name" (guard-process-kill.sh)
content/docs/releases/	AGENTS.md:679 — Documentation Guardrails row, content/docs/releases/ · RELEASE-OWNED · "❌ Never edit in a code PR." Release notes are written centrally at release time… your PR's input is its changeset

Nothing was moved into AGENTS.md: all five sentences were already there, each strictly longer
than the digest's version of it. AGENTS.md is not in this diff.

The card's premise for the gate change is FALSIFIED — measured, not argued

The card and its dispatch both state that scripts/pm/check-governed-prose.mjs "counts
CLAUDE.md as one of two instruction surfaces that must name all five governed surfaces", so a
pointer file would red it and the gate must adapt. It does not read CLAUDE.md at all.

Its PROSE_SURFACES holds AGENTS.md (Prime Directive #14) and
.claude/skills/pm-dispatch/SKILL.md. CLAUDE.md appears in that gate only as a register
entry — one of the five surfaces those two files must NAME. The gate's own verdict line says
so, and it is byte-identical before and after this PR:

✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces (docs/adr/** · .claude/** · skills/** · AGENTS.md · CLAUDE.md) and claim no others.

Three ablation legs on the committed tree, each self-restoring, each with its on-disk landing
proven before the gate ran and git diff HEAD proven empty after:

leg	mutation (on-disk proof)	gate exit	reading
B — positive control	drop `skills/**` from the AGENTS.md PD#14 region (occurrences 2 → 1)	1	the gate is live in this tree and CAN red: "AGENTS.md:255-289 … does not name skills/** — the register governs it and this prose under-claims"
A1 — the predicted red	CLAUDE.md replaced by prose with no pointer (occurrences of the string AGENTS.md in it: 0)	0	the ruling predicted red. It is green.
A2 — the decisive one	CLAUDE.md deleted entirely	0	the gate treats an unreadable input as "cannot read — Red, not a skip (#4690)", so a green run with the file absent proves it never opens CLAUDE.md

So the ruled adaptation has nothing to attach to, and the ruled self-test pair cannot be
written: "pointer form ⇒ green" and "prose without a pointer ⇒ red" are the same run of a gate
that reads neither file.

Worse, taking the ruling literally would BREAK a gate. scripts/pm/dispatch-gates.mjs pins the
separation, with a comment saying exactly what to do when a premise like this one is wrong:

t('a one-root declaration does not reach the other root file', !proseHints.some((h) => hintCovers(h, 'CLAUDE.md')));

Read from the real gates, not fixtures: what is pinned is that the tree still HAS the
declarations. If one of these gates stops reading its root file, delete its case with the
declaration — never keep a case green by re-pointing it at a gate that never read the file.

⇒ scripts/pm/check-governed-prose.mjs is not in this diff. The ruling's intent holds
untouched and unweakened: the five-surface enumeration stays enforced on the two files that
actually carry it, and CLAUDE.md stays a registered governed surface both of them must name.

The one gate that DID need this PR: the line ratchet

scripts/pm/check-skill-line-ratchet.mjs pins CLAUDE.md at a ceiling of 41 with headroom 0.
Shrinking the file to 5 leaves 36 lines of slack. The gate is green either way — its slack nudge
only fires past 120 lines — but 36 lines of budget on the one file that existed to inline
must-never-miss rules is precisely what a future digest re-accretes into. Lowered 41 → 5, the
ratchet's own convention ("a ceiling may be LOWERED by any PR that shrinks its file — lowering is
always legitimate and encouraged", headroom 0 on every entry). This is a LOWERING and needs no
ruling of its own; the maintainer ruling is quoted in the entry because it is what emptied the
file.

✓ check-skill-line-ratchet: CLAUDE.md is 41 lines (ceiling 41; headroom 0).     ← before
✓ check-skill-line-ratchet: CLAUDE.md is 5 lines (ceiling 5; headroom 0).       ← after

Every other reader of CLAUDE.md, measured

grep -rn CLAUDE.md scripts .github .claude — no gate, hook or workflow requires prose in it:

• scripts/pm/check-governed-merges.mjs — the governed register. Keeps listing CLAUDE.md, as required.
• scripts/check-required-contexts.mjs — CLAUDE.md is in the derivation population but not in INSTRUCTION_SURFACES; it enters that set "by naming [a registered context], never by sitting next to one that does", and it names none, before or after.
• scripts/check-agent-test-spelling.mjs — corpus membership only (LOOSE_FILE_NAMES); scans for test-command spellings, of which the new file has none.
• scripts/ci/select-gate-families.sh — maps the path to the agent-config family; unaffected.
• .claude/hooks/guard-main-checkout-bash.sh, check-governed-queue-guard.mjs, check-doc-frontmatter.mjs — comments and self-test fixtures only.

Governed terminal

$ node scripts/pm/check-governed-merges.mjs --test CLAUDE.md scripts/pm/check-skill-line-ratchet.mjs
governed-surface predicate: 1 of 2 path(s) hit the register (5 surfaces, repo-agnostic).
  ⛔  GOVERNED — a human merge is the review record for this PR (#9495 regime).
      No seat flips it ready, enqueues it, or arms auto-merge (AGENTS.md Prime Directive #14).
      One hit governs the whole PR — 「混合 diff 一条命中即整 PR 分叉」; proportion is not a question.
      CLAUDE.md ×1 — the repo-root Claude instruction file
  paths not on the register: scripts/pm/check-skill-line-ratchet.mjs
exit=3

Draft PR, maintainer merge. skip-changeset: nothing published moves — CLAUDE.md is a
repo-root instruction file and scripts/pm/** is repo tooling; neither is in any package's
files[].

维护者速读(草稿)

改了什么 — 根 CLAUDE.md 从 41 行缩到 5 行,只剩一句话:AGENTS.md 是唯一真相源,动手前先读它。
原先内联的四条「绝不能漏」规则(认领优先、worktree-first、禁 git stash / 禁按名杀进程、代码 PR 不
改 release notes)全部离开该文件——它们本来就一字不落地住在 AGENTS.md,本 PR 逐条核验过。
.claude/hooks/** 一行未动,三个守卫照常拦截。附带把行数棘轮里 CLAUDE.md 的上限从 41 降到 5。

为什么改 — 维护者裁决:「claude.md 是不是直接让他阅读 agents.md 即可,没必要维护两套」。两套指令
集的代价是漂移:摘要版会落后于 AGENTS.md,而只读 CLAUDE.md 的 agent 拿到的就是那份落后的摘要。
指针不会落后。

风险与代价(含回滚) — 已知代价(程序卡 #17161 已记录一次):Claude Code 会话启动自动加载
CLAUDE.md,AGENTS.md 只能经指针到达,所以跳过指针的会话在第一个 hook 触发前手里没有规则。托底的
是 hooks——worktree、stash、杀进程三条最贵的规则由 PreToolUse 守卫机械强制,不依赖任何一个 agent 读没
读文件。回滚成本为零:两个文件、一次 revert,无生成物、无 changeset、无发布面。

席位意见 — (留空,待评审席位定稿)

你要做的 — 只需确认两件事,然后人工合并这个 draft PR:① 那句指针的措辞是否是你要的口吻;②
本 PR 没有按裁决第 4 条去改 check-governed-prose.mjs——因为实测证明那条裁决的前提不成立(该门禁根本
不读 CLAUDE.md,删掉整个文件它照样绿),而照裁决字面去改会让 check:pm-dispatch-gates 变红。证据在
上面的消融表里。如果你要的是另一种处置,说一声。

Acceptance notes

Gates — all green at 41329812, the final commit.

✓ dispatch-gates --ran: 33 derived famil(ies) accounted for — 33 run, 0 NOT-MEASURED.
✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces … and claim no others.   (identical before and after)
✓ check-skill-line-ratchet: CLAUDE.md is 5 lines (ceiling 5; headroom 0).
✓ dispatch-gates self-test: 1624 cases pass.        (pnpm check:pm-dispatch-gates, exit 0)
$ pnpm lint            # eslint . --no-inline-config, repo-wide, exit 0

Re-derived on the actual diff with
node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths) at
merge commit 41329812: 33 commands, all run, all exit 0. The dispatch's 34-command lead
included pnpm check:pm-governed-prose; it drops out because the re-derivation sees the real
diff, which does not touch that gate. Ran it anyway, before and after — green both times, same
verdict line. pnpm check:pm-dispatch-gates was run detached with its exit captured to a file
and waited on in the foreground, per the dispatch.

Shallow checkout. git rev-parse --is-shallow-repository is true here. No ancestry verdict
in this PR rests on a negative merge-base --is-ancestor; the AGENTS.md survival checks are
content greps on the working tree at origin/main, which a shallow checkout answers exactly.

origin/main merged in. f801e7d7 (#17142) landed mid-run and touched
scripts/pm/dispatch-gates.mjs, the derivation tool itself, which the tool reported as a STALE
TREE warning. Merged and re-derived on the current tool: the 33-command list is byte-identical to
the pre-merge one.

Out of scope, noted, not filed. Six script/agent comments cite CLAUDE.md as the place a
rule is stated — scripts/check-vendor-version-stamps.mjs:226, scripts/check-docs-single-h1.mjs:109
and :117, scripts/import-prerequisite.mjs:12 and :382, scripts/pm/ci-failure.mjs:215,
.claude/agents/os-dev.md:35, .claude/workflows/docs-accuracy-audit.js:41. After this PR those
rules are stated in AGENTS.md instead. Not filed and not fixed here: CLAUDE.md still exists
and points at AGENTS.md, so every one of those citations still routes a reader to the rule —
one hop longer, never a dead end. Who picks this up: the sibling sub-issues of #17161 do the same
rewrite in objectui and cloud, and whoever lands the last of them is the natural place to
re-point the citations across all repos in one pass. No gate reads them.

---

Generated by Claude Code

---

Generated by Claude Code

[os-justin]

维护者速读 — PR #17168(#17162,skills 席终稿,2026-09-09T14:28Z)

改了什么 — objectstack 根 CLAUDE.md 从 41 行缩到 5 行:标题加一句「AGENTS.md 是唯一事实源,第一次改文件前通读;一套指令只维护一处,这里不复述任何规则、钩子清单或守卫,以后也不加」。原先内联的四条「绝不能漏」规则(认领优先、worktree-first、禁 git stash / 禁按名杀进程、代码 PR 不改 release notes)全部离开该文件——五句原话都在 AGENTS.md(:217 / :318 / :393 / :528 / :679),本席逐句核过。.claude/hooks/** 一字未动。附带把行数棘轮里 CLAUDE.md 的上限从 41 降到 5(降上限不需裁决,注释里引了你的原话)。这是你「claude.md 直接让他阅读 agents.md 即可,没必要维护两套」在 objectstack 的落地,程序父单 #17161。

为什么改 — 两套指令集的代价是漂移:摘要落后于原文,而只读 CLAUDE.md 的 agent 看不出它过期;指针没有可过期的东西。

风险与代价(含回滚) — 已知代价:Claude Code 会话自动加载的是 CLAUDE.md,AGENTS.md 经指针到达;托底是三个 PreToolUse 钩子原样在。回滚 = revert 两文件一提交。

席位意见 — 接受。本席在分支头复核:五句原话在;check-governed-prose 前后逐字节相同的绿(本席派发词里让 dev 改这个门禁那一条是本席读错了——该门禁根本不读 CLAUDE.md,dev 用三条消融证明后拒绝执行,是对的,本席已在 ACCEPT 里认下);棘轮本席重跑绿(CLAUDE.md is 5 lines (ceiling 5; headroom 0));受管判据 exit 3;merge-tree 干净;CI 19 绿 / 10 跳过 / 1 在跑(14:27Z,Lint & Repo Gates)。dev 顺手发现的八处「规则见 CLAUDE.md」注释引用(多一跳、不断链、无门禁读它)已记到 #17161,由程序最后一个子单一次性改指向。

你要做的 — 看一眼那五行的口吻;CI 全绿后人工合并,或由 os-zhuang / hotlong 批准后本席入队(受管面,席位不翻 ready、不入队、不挂 auto-merge)。

---

Generated by Claude Code