[os-justin]

Fixes #18636

Clause-②: no

The defect

The round-open step-2 reading at .claude/skills/pm-dispatch/SKILL.md:91 named three paths — .claude/{settings.json,agents/*.md,hooks/*} — while the harness loads the whole .claude/** tree from the primary checkout at session start and never reloads it, .claude/skills/** in full for every seat session (the load moment scripts/pm/check-skill-line-ratchet.mjs's header states). So the charter told the seat to measure a set SMALLER than the set that rots it. The tool half — scripts/pm/check-harness-current.mjs's PATHS widened to .claude/skills/**, plus a second CONTENT reading per path — landed in PR #18610; this PR is the charter half that #18544 comment 5706565860 said stays on this file, line-neutral. The card's own true positive: SKILL.md moved ab1d359 → 7f4aa9e5c at a round open and the step-2 reading said nothing, because the seat's copy of the tool was the old one too. The charter has to name the right set on its own, whichever copy of the tool the seat holds.

The line, before and after

:91 only; nothing else in the file, no reference file, no script. Before (120 B, on origin/main 99fcb4ac1, byte-identical to the cf39b83c0 the card was dispatched against):

• 同读 harness 载入面 .claude/{settings.json,agents/*.md,hooks/*} 的最新触碰是否已在共享检出 HEAD。

After (96 B):

• 同读 harness 载入面 .claude/** 的最新触碰与内容是否已在共享检出 HEAD。

Payment: in place, line-neutral — 812 / 812, 1 insertion / 1 deletion (git diff HEAD~1 HEAD --stat), no second line, no re-wrap, no other line moved, the widest table row unchanged (342 B, pin 342). It reads as one block with :92, which is untouched: 「否 ⇒ 记座位贴,自然收班时换会话,⛔ 不中断批次;读数走 scripts/pm/check-harness-current.mjs。」 — the 「否」 still answers :91's yes/no question, and :92 still names the tool that takes the reading.

Two things moved on the line:

• .claude/{settings.json,agents/*.md,hooks/*} → .claude/** — the SET, the card's fix.
• 「最新触碰」 → 「最新触碰与内容」 — the READINGS. The tool on main takes two per watched path (PLACEMENT: is the commit that last touched the path on origin/main an ancestor of the shared HEAD; CONTENT: are the shared HEAD's bytes of the path equal to origin/main's), and a path is STALE on either. The charter now names both, so a seat holding an old copy of the tool — the card's reading 3 — still knows what it is supposed to measure and can see when the tool's report is short of it.

Why the set is the directory and not a longer list — measured, then the four axes

Measured on this tree at 14f751515 (git ls-files per pathspec): .claude/** selects 43 tracked files; the tool's four pathspecs select 1 (settings.json) + 1 (agents/*.md) + 10 (hooks/*) + 29 (skills/**) = 41. The two files outside the four pathspecs are .claude/launch.json and .claude/workflows/docs-accuracy-audit.js (both last touched 5de93728a; their load moment is unmeasured, as the tool's header says of hooks and agents). Spelling the four pathspecs on the line would cost 131 B — over the 120-byte max-line budget — and would be the file-list spelling the tool's own PATHS comment names as the failure shape: 「naming the charters one by one is how a predicate ends up guarding less than it claims (the population it covers must be the population that is loaded)」.

四维分析(新记录,按章程用中文):

• 实际业务需求 — 场景已实测,不是推断:[finding] check-harness-current's PATHS omits .claude/skills/**, so a seat seating on a stale checkout is never told its charter is stale #18544 评论 5706565860 记录了一个席位整班执行已离开 origin/main 的旧 :92 处方(车道停摆约 7 小时);本卡读数 2 是同一机制的第二次真阳性(SKILL.md ab1d359 → 7f4aa9e5c,step-2 读数一字未提)。harness 侧的载入面本来就是目录——check-skill-line-ratchet.mjs 头部「read in full by every seat session」说的是 .claude/skills/**,不是三个文件;写 .claude/** 就是写实测的载入面。
• 项目长远合理性 — 契约优先:章程是契约,工具是读数的实现,章程应当说清「量哪个集合、量什么」,由工具追章程而不是章程追工具的常量。工具头部自己写明「the population it covers must be the population that is loaded」。逐个列 pathspec 的写法今天恰好等于 PATHS,.claude/ 明天多一个子目录就再次落后——正是本卡的缺陷形态,只是换了个数字;目录写法不随 PATHS 腐烂。
• 防 AI 写代码犯错,尤其是防 AI 写元数据 app 犯错 — 席位是 AI,它量什么集合只取决于章程这一行字。读数 3 已证工具副本也会旧:章程点名 .claude/** 与两种读数之后,旧工具只报两条路径时席位能在结构上看出集合不齐;现在的三路径写法让它无从察觉。不写「与上轮」:那是一条形态未定的席位侧账本义务,120 B 里写不清的义务恰是 AI 会写错的那种。
• 创业阶段不扩散需求 — 零新机制:一行、两个词,两种读数都已在 main 的工具里落地;不加参考文件、不加脚本、不加座位贴字段。目录写法多覆盖的两个文件(launch.json、workflows/*.js)载入时刻未测,按从紧原则不据此扩 PATHS,记入验收备注交工具所在车道。

The directory-diff suggestion — taken in one half, left in the other

The card proposes writing 「目录 diff 与上轮逐字节比对」 into step 2. Split along the same axes:

• Taken — the byte-level half, as the tool's CONTENT reading against origin/main, which PR fix(pm): check-harness-current watches the skills tree, and compares content as well as placement #18610 landed and which :91 now names (「与内容」). It is strictly more informative than the seat's byte-count habit the card describes: it names the differing files per path, needs no history, and fires on a checkout that was ALREADY stale at seating.
• Left — the 「与上轮」 half (a number compared with the previous round's). (i) For the three charters, :89–:90 already compare the latest touch sha with the previous round-open marker and prescribe the re-read, so a second round-over-round comparison on :91 would duplicate them. (ii) For the harness-loaded rest (settings.json, hooks, agents) a re-read is not a remedy — the harness never reloads — so :92 is the whole prescription, and "moved this round" versus "already stale" changes nothing the seat does. (iii) A round-over-round ledger is a new seat-side obligation whose form (what to record, where, compared how) does not fit in the 24 bytes the line has left, and an under-specified obligation is the kind an AI seat gets wrong. Where it would go if the seat wants it: the round-open marker's contents on :90 (record the tool's named stale paths beside the touch sha), or a since-marker flag on the tool — not :91.

Ratchet reading

pnpm check:pm-skill-ratchet (self-test then gate) at 14f751515, exit 0:

✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md: widest table row is 342 bytes (pin 342; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 812 lines (ceiling 812; headroom 0).

The edited line went 120 B → 96 B, under the 120-byte max-line budget; .claude/skills/pm-dispatch/SKILL.md is the only path in the diff; grep -naP for control bytes over the file: 0 hits.

Gates

Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack from the worktree with no hand-fed paths (change set from git: 1 path vs merge base 99fcb4ac1, three-dot): 18 commands. Each run with the tool's own capture idiom (redirect, then $?, the printed command recorded byte for byte), then reconciled with --ran: 18 derived, 18 run, 0 NOT-MEASURED, 0 UNRUN (exit 0, 「a DERIVED zero — all 18 recorded an exit code and none of them is 3」).

#	command	exit
1	node scripts/check-closing-keyword-parity.mjs	0
2	node scripts/check-closing-keyword-parity.mjs --self-test	0
3	node scripts/check-comment-mask-corpus.mjs	0
4	node scripts/pm/check-governed-queue-guard.mjs --self-test	0
5	node scripts/pm/check-harness-current.mjs --self-test	0 (26 cases)
6	pnpm --filter @objectstack/lint run check:doc-formula-expressions	0 — first run exit 3 PREREQUISITE NOT MET (unbuilt closure), not a measurement; pnpm exec turbo run build --filter=@objectstack/lint under scripts/pm/os-verify-lock.sh (VERDICT command-exit 0, held 141 s, waited 0 s), rerun exit 0
7	pnpm check:agent-test-spelling	0
8	pnpm check:doc-authoring	0
9	pnpm check:driver-memory-census	0
10	pnpm check:nul-bytes	0
11	pnpm check:pm-governed-merges	0
12	pnpm check:pm-governed-prose	0 (2 surfaces name all 5 governed surfaces)
13	pnpm check:pm-half-states	0
14	pnpm check:pm-skill-id-lint	0 (27 files clean)
15	pnpm check:pm-skill-ratchet	0
16	pnpm check:refd-timer-probe	0
17	pnpm check:skill-frame-sync	0 (one declared copy, 4 axes, 74 markdown files scanned)
18	pnpm check:watch-hint-literal	0

Not derived, run anyway:

• node scripts/pm/check-governed-merges.mjs --test .claude/skills/pm-dispatch/SKILL.md → exit 3, GOVERNED (1 of 1 path hits the register) — the expected answer for this path, recorded as such and not as a failure.
• pnpm check:pm-settings-deny-roster (the derivation flags its roster as living under .claude, so its silence is evidence in neither direction) → exit 0, 17 declared = 17 enforced.
• node scripts/check-skill-frame-freshness.mjs --no-fetch → exit 0.
• Tier: dispatch-gates.mjs --tier --repo objectstack-ai/objectstack .claude/skills/pm-dispatch/SKILL.md → the MANDATORY tier the claim comment's Container & model line carries, derived from the file surface; this run was at that tier.

pnpm lint (the repo-wide union) is CI's run; locally it narrows to zero files by measurement, not by skipping: (1) population read from eslint.config.mjs itself — every files: glob is {ts,tsx,mts,cts,js,jsx,mjs,cjs}, no markdown; (2) npx eslint --no-inline-config --format json on the one changed path → 1 file, 「File ignored because no matching configuration was supplied」, 0 errors, exit 0; (3) invariance: the diff touches no file eslint lints, so no untouched file's verdict can move. Not measured locally: the 52 artifact-roster families, the 11 wide-population families and the 14 pending-changeset families the derivation lists outside its runnable total — CI's, as the derivation says.

Changeset

skip-changeset, measured on the fast track: .claude/** is repo-internal agent tooling, in no package's files[]; nothing published moves.

Acceptance notes

• noted, not filed: on 14f751515 the tool's four PATHS pathspecs select 41 of the 43 tracked files under .claude/; the two outside are .claude/launch.json and .claude/workflows/docs-accuracy-audit.js. Whether the harness loads either at session start is unmeasured (the same status the tool's header gives hooks and agents), so this is an observation, not a contract violation; 承接者: the tool's lane (domain:devx, scripts/pm/check-harness-current.mjs), if a load-moment measurement ever lands.
• noted, not filed: check-doc-formula-expressions refuses with exit 3 on a fresh worktree until @objectstack/lint's closure is built; its refusal text names the exact fix and the derivation classifies 3 as NOT-MEASURED, so nothing is hidden — 承接者:无.

维护者速读(草稿)

• 改了什么:.claude/skills/pm-dispatch/SKILL.md 第 91 行一行。开轮 step-2 的 harness 读数原来只点名三个路径 .claude/{settings.json,agents/*.md,hooks/*},现在点名整个载入面 .claude/**,并把工具在 main 上已落地的两种读数(最新触碰、内容)一起写进去。812 / 812 行不变,该行 120 B → 96 B,别处一字未动。
• 为什么改:harness 在会话开始时整份读入 .claude/**(.claude/skills/** 每个席位都整份读),章程却让席位只量三个文件——量的集合比会腐烂的集合小。[finding] check-harness-current's PATHS omits .claude/skills/**, so a seat seating on a stale checkout is never told its charter is stale #18544 评论 5706565860 记录过一个席位整班执行旧规则、车道停摆约 7 小时;本卡又撞上一次(SKILL.md 动了,读数一字未提,因为席位手里的工具也是旧的)。工具那一半 PR fix(pm): check-harness-current watches the skills tree, and compares content as well as placement #18610 已合;这是章程那一半。
• 风险与代价(含回滚):风险极低——一行文字,不加机制、不加文件、不动棘轮上限;18 个派生门禁全绿,棘轮 812 / 812、最宽行 342 不变。回滚就是 revert 这一个 commit(14f751515),不牵连任何脚本。
• 席位意见:(留给席位)
• 你要做的(一个动作):批准或合并这个 draft PR(受管面 .claude/**,须维护者亲手或授权批准后由席位落地)。

---

Generated by Claude Code

[os-justin]

维护者速读(终稿)

skills 席 session_01Gqi43smmqjJ5sUrhfoPeKu,2026-09-17T12:56Z。对照本席自己读的 diff 校正 dev 草稿;席位意见已填。

改了什么 — pm-dispatch/SKILL.md 开轮第 2 步的 harness 读数一行(:91)。原文点名三个路径「.claude/{settings.json,agents/*.md,hooks/*} 的最新触碰」,改为「.claude/** 的最新触碰与内容」。只此一行,812 / 812 行不变,120 B → 96 B,其余文件一字未动。

为什么改 — harness 实际载入的是整个 .claude/** 目录,.claude/skills/** 每个席位会话整份读;而章程教席位去量的集合比真正会让它腐烂的集合小。#18544 已经撞上真阳性:本章程自己从 ab1d359 动到 7f4aa9e5c,而按旧读数跑的工具一个字没说。工具那一半(check-harness-current.mjs 的 PATHS 纳入 .claude/skills/**)已由 PR #18610 落地;本 PR 是章程这一半,把集合写成与工具一致的载入面本身。「与内容」点的是工具的第二个读数(内容对照 origin/main),正是抓到那次真阳性的读数。

风险与代价(含回滚) — 规则层受管面,只改一行定义;不加标签、不加义务、不动工具。卡上的第二条建议(逐轮目录字节比对)没有采纳::89–:90 已经对照上一开轮标记比对三章程的触碰,而 harness 载入的其余部分重读无用(:92 是全部处方),逐轮账本是一项塞不进这一行的新义务;若要,落点在 :90 的标记内容或工具的 since-marker 开关。回滚 = revert 一个提交,零副作用。

席位意见 — 荐合。它把章程与已落地的工具对齐,消除的是一个已经发生过的漏读;不新增规则。

你要做的(一个动作) — 认可「集合 = 载入面本身」即 Approve 本 PR(或直接人工合并);席位随后按裁定 C 落地。若你想要逐轮目录比对也进章程,Request changes,本席另起。

---

Generated by Claude Code