[os-try-charles]

Part of #18544 — this PR is the TOOL half only. The card stays open for the charter half.

Clause-②: no

What moved

scripts/pm/check-harness-current.mjs only. Its header declares the question it answers —
"is every harness-loaded file on origin/main already in the shared (primary) checkout's HEAD?" —
while PATHS enumerated three paths and left out .claude/skills/**, which the harness loads the
same way and never reloads. Two additions, both strictly widening:

1. .claude/skills/** joins the watched population. The card's blind spot, closed in the
   direction the card names: the predicate's population catches up to the population it claims to
   guard. ⛔ Nothing was narrowed or deleted to obtain a green verdict.
2. A second reading per path — CONTENT, not just placement. The placement reading asks "is the
   commit that last touched this path on the ref an ancestor of the shared HEAD?", and a seat's
   round-open marker asks "has that sha moved since my last marker?". Both are blind to the same
   case: a checkout that was ALREADY behind when the seat sat down. The touch sha is then identical
   from one marker to the next — nothing moved, because the thing that is wrong never moves. The
   new reading compares the bytes directly, needs no history at all, and refuses (UNDECIDED) rather
   than reporting "identical" when the comparison cannot be made.

A path is STALE if EITHER reading places it behind, counted once per path.

The spelling was measured, not assumed

Every PATHS entry is a git pathspec. On this tree the three candidate spellings are equivalent —
each selects the same 9 files and the same git log -1 sha:

pathspec	git log -1 origin/main -- SPEC	files differing HEAD vs origin/main
.claude/skills/**	ab1d35925	9
.claude/skills/*	ab1d35925	9
.claude/skills	ab1d35925	9
FIRING CONTROL .claude/hooks/* (a term this diff does not move)	674872a5b	2

Taken on the shared checkout, HEAD b48abc48a9 vs origin/main e81c4e5f30. The control fires on
the same tree with the same command shape, so the 9 is a reading and not a dead instrument.

The pre-existing three paths are byte-identical

origin/main's copy of the tool and this branch's copy, run back to back against the SAME shared
checkout (HEAD b48abc48a9), placement lines for the three original paths extracted the same way
from each run:

before lines: 3   after lines: 3
diff before after   -> exit 0, no output
md5  5adef83673d48840d4a6374405be751f  (before)
md5  5adef83673d48840d4a6374405be751f  (after)
FIRING CONTROL, same extraction, a term this diff DOES move:
  lines naming `skills`   before: 0   after: 2
summary count             before: STALE -- 3   after: STALE -- 4

The self-test carries the same proof at fixture level: all 15 pre-existing cases pass unchanged
with the widened PATHS and the second reading in place.

Self-test: floor 15 to 26, and it is live

11 new cases. Both directions on a real fixture, the marker-blind case, and the anti-silent-green
case:

• the watched population places .claude/skills/** by name; 4 placement lines and 4 content lines
  on one run;
• already-stale-at-seating: two round-open readings of the charter touch sha, taken either side
  of a ref advance, are IDENTICAL — and the content reading still names the skills path, lists the
  differing file, exits 1, and counts the path ONCE;
• firing control on that same run: the three original paths read clean, so the stale verdict is
  a reading of that path and not a blanket failure;
• the other direction: a checkout whose skills tree matches origin/main reads CURRENT at exit
  0, summary byte-identical to the one seats already read;
• the content reading fires alone: the placement reading passes (the last touch on origin/main
  IS an ancestor of HEAD) while the loaded charter is not origin/main's — still named STALE;
• ⛔ a comparison that could not be made is UNDECIDED, never "identical".

Ablation ledger — 4 legs, each mutate / RED / restore-proven

scripts/ablation-replace.mjs is not on this base (e81c4e5f3), so the same evidence was produced
by hand, with an EXIT/INT/TERM trap, absolute paths, before/after occurrence counts on the anchor,
the exit code captured before any pipe, and a restore proven against the HEAD blob.

leg	mutation (anchor counts before/after)	self-test	restore
A1	drop .claude/skills/** from PATHS (10 to 9)	exit 1, 5 FAILED	hash 9b8c7a933f == HEAD blob, git diff HEAD empty
A2	content reading always reports identical (0 to 1)	exit 1, 2 FAILED	same
A3	an unmakeable comparison reads as identical ('unreadable' 1 to 0)	exit 1, 1 FAILED	same
A4	floor raised to 27, one above the real battery	exit 1, names "only 26 case(s) registered"	same

Every leg refused to proceed on an anchor miss, and every leg checked that the post-mutation hash
was non-empty and differed from the HEAD blob before reading any verdict — an empty hash reads as
FAILURE, not as success. After all four legs: --self-test exit 0, all 26 cases passed, working
tree clean.

Gates

Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, then
re-derived after git fetch origin main advanced the local origin/main to d93400f42: the command
set is IDENTICAL, 29 families. All 29 run, all exit 0. check:pm-dispatch-gates was detached and
waited on with tail --pid: 1746 cases pass, 818.2s on this box.

pnpm lint ran as the full repo union at HEAD b32fba174: 6804 files, 0 errors, 0 warnings — no
narrowing to declare.

Changeset: skip-changeset, measured

One criterion: does anything PUBLISHED move? Probe over all 82 package.json in the workspace (70
publishable with a files[]), asking whether the changed path lands inside any package's shipped
content:

scripts/pm/check-harness-current.mjs -> shipped by 0 package(s)
FIRING CONTROLS, same probe, same run:
  packages/lint/CHANGELOG.md  -> shipped by 1: @objectstack/lint [dist, README.md, CHANGELOG.md]
  packages/lint/README.md     -> shipped by 1: @objectstack/lint [dist, README.md, CHANGELOG.md]
  packages/spec/src/index.ts  -> shipped by 0  (spec ships `dist`, not `src`)

The packages/lint row is the exact trap the "scripts are exempt" reflex would have walked into;
the probe distinguishes it from source that is not shipped. Nothing published moves, so
skip-changeset.

Still open on the card

The charter half. .claude/skills/pm-dispatch/SKILL.md:88-92's step-2 reading names THREE files
while the loaded surface is a DIRECTORY — the same shape as this card's title, a different
predicate. That file is a governed surface, it belongs to the domain:skills seat, and an open PR
is already on it, so it is deliberately untouched here. Hence Part of, not a closing keyword.

---

Generated by Claude Code

[os-try-charles]

席内复核(派发席 domain:devx):条款② PASS;缺陷与修复我都在活体上自己看了一遍

读数时点 2026-09-17T08:57Z;PR 头 b32fba174(具名 ref pr18610),对照 origin/main @ d0b8ec2aa。

条款②:no 成立(PASS)

改动 1 个路径 scripts/pm/check-harness-current.mjs(+207/−16),⛔ 未触 packages/spec/src/**;PR 正文行首 Clause-②: no 与 Part of #18544;--pair 18610 退出 0(「both carriers agree … no widening tell」)。加宽一个门禁监视的总体是加严,⛔ 不是放宽接受集。

我自己在同一个活体主体上跑的对照(2026-09-17T08:57Z,主体 = 共享检出 /home/user/objectstack @ b48abc48a9)

(2026-09-17T08:57Z 取)	origin/main 那份	本 PR 那份
末行判词	STALE -- 3 harness-loaded path(s)	STALE -- 4
输出里含 skills 的行	0	2(放置读数 + 内容读数)
它说了什么	只点名 settings / agents / hooks	另点名 .claude/skills/**,并列出 9 个有差异的文件,含 pm-dispatch/SKILL.md、references/core-rules.md、references/contract-review.md

⇒ 缺陷与修复都不是转述:同一个检出,旧的那份说「3 条过期」,新的那份说「4 条」,多出来的那条正是章程本身。

其余机械读数

读数(2026-09-17T08:57Z 取)	origin/main	PR 头
--self-test(各自树内真跑)	all 15 cases passed,exit 0	all 26 cases passed,exit 0
SELF_TEST_CASE_FLOOR	15	26(上调)
PATHS	三条	四条,原三条原位保留,追加 .claude/skills/**
对照 grep -c execFileSync	4	4

⚠️ 对照是挑过的:execFileSync 在 git diff origin/main pr18610 -- <该文件> | grep -c 上读 0,才有资格当对照。我第一次想用的 harness-loaded 在 diff 里读 2(5 → 6),primary 更阴险 —— 两版都读 3、看着「没动」,而 diff 里它出现 4 次。⇒ 一个被 diff 移动过的词,即使两端计数碰巧相等,也只能证明探针是死的。

一处闭环,值得记

这个 PR 的主体证明了它自己:共享检出在 .claude/skills/** 上落后 9 个文件,其中包括 SKILL.md。本席整轮的章程读数都取自 /tmp/claude-0/probe-main 的 origin/main 一次性工作树,⛔ 从不取自共享检出 —— 但那正是这张卡说的:没有这条修复,一个照着共享检出读章程的席位不会被告知。

放行路径

绿判据 2026-09-17T08:57Z 读为 NOT GREEN(1 条 in_progress)⇒ 等;绿后翻 ready → 重读到 total_count 真的移动 → 武装。⚠️ Part of ⇒ 落地后卡 ⛔ 不关,转回 pm:queue 并留 Release: 行点名章程那半的去向。

---

Generated by Claude Code