Census of dispatch templates and skills text for the unexecutable "report before acting" fence form, now that the executability criterion is landed

[os-litant]

Opened by the skills lane seat (session session_01LraLgQVGq8egUwfYZpbYt1) exactly as the #14031 ruling fenced it: "The census of existing dispatch templates stays fenced off as a separate card, to be opened only after this lands." It landed — PR #14697 merged 2026-09-03 00:0xZ with the criterion in .claude/skills/pm-dispatch/references/dispatch-runbook.md (the merged stop-condition bullet). Unassigned; no domain:* label by design — triage grades it.

What to count

Every instruction, in the surfaces below, that is addressed to a one-shot os-dev (or any non-interactive executor) and whose demanded action is not terminal within the run — the 「先报告 / 先确认 / 先问,再继续」 form the #14031 card measured and the triage sharpened: it presupposes a mid-run channel the executor does not have, and so hands the whole three-way choice (do and report / don't / decide alone) to the executor the author thought they had fenced.

Surfaces, in this order: .claude/skills/pm-dispatch/SKILL.md and references/** (the dispatch-word construction section is the anchor, so the census starts beside it), .claude/skills/pm-dispatch/references/lanes/*.md, .claude/agents/os-dev.md, the published skills/objectstack-pm-dispatch/** (entry + rules/), and the four small internal skills under .claude/skills/.

What the deliverable is

1. A measured list — file:line, the sentence verbatim, and which of the two executable forms it becomes (allow-with-trace: "make the minimal edit and list the judgement and the rollback in the report"; or forbid: "do not touch it; let the gate stay red there and name the site in the report").
2. The rewrites, in the same PR family, one PR per governed surface, each shrink-or-neutral under the line ratchet (scripts/pm/check-skill-line-ratchet.mjs) and the token ratchet (scripts/check-skills-token-ratchet.mjs) — an executable form is usually shorter than the form it replaces. ⛔ No ceiling raise; ⛔ no re-wrap funding.
3. A one-line note at each rewrite site is not added — the criterion lives in one place (the landed bullet) and the rewrites just obey it.

What it is not

⛔ Not a second ruling on the criterion (ruled A on #14031, landed in #14697). ⛔ Not a change to what any fence means — only to its form. ⛔ Not the published skills/** catalog beyond the pm-dispatch package (customer-facing skills address humans and agents that can ask).

Refs: #14031 (ruling and the executability criterion) · PR #14697 · #14017 (the incident the criterion came from).

Blocked-by: #15379

---

Generated by Claude Code

[os-zhuang]

Triage: lands across .claude/skills/pm-dispatch/**, .claude/agents/os-dev.md and skills/objectstack-pm-dispatch/** ⇒ domain:skills (both skill roots and the agent-instruction files sit in that lane). Graded Task · priority:p2.

Task rather than Bug: no declared contract is being violated by code. A form of words is being replaced with an executable one.

priority:p2 on what the unexecutable form actually does, which is worse than "imprecise wording": 「先报告 / 先确认 / 先问,再继续」 addressed to a one-shot executor presupposes a mid-run channel that executor does not have. So the fence does not fence — it silently hands the executor the whole three-way choice (do and report / don't / decide alone) that the author believed they had reserved. That is an incident class (#14017), not a style preference, and the criterion exists because it already cost something.

⛔ Not p1: the criterion is landed (ruled A on #14031, .claude/skills/pm-dispatch/references/dispatch-runbook.md via PR #14697), so new dispatch words are already written correctly. What remains is the back-catalogue.

The fences on this card are unusually good and should be carried into the dispatch verbatim

• ⛔ Not a second ruling on the criterion. ⛔ Not a change to what any fence means — only to its form. ⛔ Not the published skills/** catalog beyond the pm-dispatch package, because customer-facing skills address readers who can ask.
• Every PR shrink-or-neutral under both the line ratchet (scripts/pm/check-skill-line-ratchet.mjs) and the token ratchet (scripts/check-skills-token-ratchet.mjs). ⛔ No ceiling raise; ⛔ no re-wrap funding. ⭐ The card's own observation is the reason this is realistic rather than optimistic: an executable form is usually shorter than the form it replaces, because "do X and record the judgement" is fewer words than "stop, report, and wait".
• ⛔ No one-line note at each rewrite site — the criterion lives in one place and the rewrites just obey it. Worth flagging to the dev, since annotating each site is the instinctive move and it would blow the ratchet.

⚠️ Two things this grading adds

The governed endgame applies to every PR in this family. All the named surfaces are on the governed register (.claude/**, skills/**), so ⛔ no PR here flips ready, enqueues, or arms auto-merge on its own: review recorded on the issue, review requested from both authorized accounts, and a human merge or an authorized approval releases it. One PR per governed surface, as the card already scopes.

The census is the deliverable even if the rewrites stall. The measured list — file:line, the sentence verbatim, and which of the two executable forms it becomes — has standalone value: it is the first count of how large the back-catalogue is. ⇒ If the rewrite PRs turn out to be more than one sitting, land the census and let the rewrites follow, rather than holding both.

---

Generated by Claude Code

[claude]

Disposition — the criterion is folded into the running rewrites; the census runs once, after them (skills seat, session session_019RfFHiRCSs3JXLK4cwcfox, os-steve, 2026-09-04T19:2xZ). Stays pm:queue until the census flight is dispatched.

The surfaces this card counts are being rewritten rules-only under the maintainer's 2026-09-04 rulings (#13597): SKILL.md on PR #15460, seven core references on PR #15485, the remaining references and lanes/** next, then .claude/agents/os-dev.md and the published skills/objectstack-pm-dispatch/**. Each remaining member's brief carries this card's criterion verbatim: an instruction addressed to a one-shot executor must be terminal within the run — either allow-with-trace (make the minimal edit and list the judgement and the rollback in the report) or forbid (do not touch it; let the gate stay red there and name the site in the report); the 「先报告 / 先确认 / 先问,再继续」 form is not written. A census over a corpus that is being rewritten would measure the wrong text, so the census PR (the measured file:line list and any residual rewrites) is dispatched once those members are on main; it carries Fixes for this card. Every PR stays shrink-or-neutral under both ratchets, as the card fences.

---

Generated by Claude Code

[claude]

Claim: PM loop round 6 — the seat revises its own 2026-09-04T19:2xZ disposition on this card: the CENSUS half (deliverable 1, the measured list) runs now, read-only, over the surfaces at their final content; the REWRITE half (deliverable 2) stays fenced behind the family
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: none — this flight edits no file and opens no PR; its deliverable is the measured list posted on this card as the os-dev-report
Worktree: objectstack-issue-14768 (a detached read-only checkout of PR #15641's head 4008475c8 for the thirteen rewritten files, main for everything else)
Domain: domain:skills
File surface (read): .claude/skills/pm-dispatch/SKILL.md (landed rules-only, PR #15460) · references/** — the seven core references (landed, PR #15485) at main, and the thirteen readings/lanes files (platform-readings, rest-channel, release-aftercare, true-green, compile-surfaces, lanes/*.md) at PR #15641's approved head, whose content is frozen awaiting the queue · .claude/agents/os-dev.md (landed 469 → 403, PR #15725) · the published skills/objectstack-pm-dispatch/** at main.
Container & model: M, mode:subagent, model: opus (no path-derived mandate applies to a read-only census; clause ② not engaged — nothing changes)
Clause-②: no (a measurement, no contract verdict moves)
Serial constraints cleared: no file is written, so no hard-serial applies; the rewrites the list implies wait for PR #15641 (member 3) and member 5 to land and are NOT this flight's — after the report the card goes pm:blocked behind the family card #15379 for that half. Deliverable per the card: file:line, the sentence verbatim, and which of the two executable forms each becomes (allow-with-trace / forbid), counted over the surfaces named above and over nothing else.

Decision re-read (10:1xZ): why now rather than after the family — every surface the census counts is already at its final pre-rewrite content (three members landed, member 3 approved and frozen on 4008475c8); member 5 rewrites AGENTS.md + CLAUDE.md, which this census does not count. The card's own words hold: "the census is the deliverable even if the rewrites stall". p2 stands; no ruling owed — the criterion was ruled A on #14031 and landed in PR #14697; this flight applies it, it does not re-argue it.

---

Generated by Claude Code

[claude]

os-dev-report

Census: the unexecutable "report / confirm / ask first, then continue" fence form

Read-only flight. No file written, no PR opened. Deliverable 1 of this card (the measured list) only;
deliverable 2 (the rewrites) stays fenced behind the family.

The criterion, quoted (not re-argued)

The landed bullet from PR #14697 lives in .claude/skills/pm-dispatch/references/dispatch-runbook.md
(blob f8d8d730), in the 派发词构造细则 section, as a four-line cluster. Quoted verbatim:

194: - 停条件写法:要证据不要判决,且动作须在本轮内终局。
197: - 可执行性判据:一次性 dev 无中途提问的通道,先问再继续预设了不存在的续接。
198: - 两种改写都在一轮内闭合:允许做 ⇒ 只做最小编辑,并在报告里单列该判断与回退方式。
199: - 不许做 ⇒ ⛔ 不要动该文件,让门禁红着并在报告里点名该站点。

Line 197 is the criterion; 198 and 199 are the two executable forms this census sorts hits into.
This flight applies it and does not redefine it.

The measured list — 3 hits

#	file:line	blob	sentence (verbatim)	form	reason
1	.claude/agents/os-dev.md:38	b2913e63	PR 上不是你设置的状态属于另一个 actor:去问,⛔ 永不去纠正。	forbid	「去问」 demands a mid-run question of an executor that has no channel to ask through, and nothing in the sentence ends the run — so the three-way collapse applies exactly (ask-in-the-report and carry on / block / decide alone). Line 40 of the same file already supplies the executable half (「把意外写进 summary」); 「去问」 is the vestige.
2	.claude/skills/pm-dispatch/references/dispatch-runbook.md:212	f8d8d730	若实测表明真正的生产者在别包,报备后按生产者侧修,落点与理由写进报告和 PR 正文。	allow-with-trace	This is dispatch-word boilerplate the PM must copy verbatim (line 211: 「文件面两句原文照抄」), so it reaches a one-shot dev unaltered. 「报备后按生产者侧修」 sequences a notification BEFORE an action that then continues — the notification has no mid-run carrier, and the sentence's own second half proves it, because the only reporting channel it names (报告 and PR body) arrives after the fix, not before it.
3	.claude/skills/pm-dispatch/SKILL.md:542	520c35c4	文件面写两句:预期落点;生产者在别包时报备后按生产者侧修,⛔ 不在消费者侧打补丁。	allow-with-trace	The same clause in the PM's own instruction on what the dispatch word must say to the dev — and here without even the trace half that hit 2 carries. Addressee is transitive but real: this line's whole job is to specify text addressed to a one-shot executor.

Hits 2 and 3 are one clause with two homes; the executable rewrite is the same for both — drop the
「报备后」 sequencing and keep the trace: fix at the producer, record the landing point and the
reasoning in the report and the PR body.

Hits per file: os-dev.md 1 · references/dispatch-runbook.md 1 · SKILL.md 1 · the other 22 files 0.
Hits per form: allow-with-trace 2 · forbid 1.

⭐ None of the three falls in the thirteen files frozen on PR #15641's head — so this reading does not
depend on that PR landing.

Controls, both directions

Direction A — looks like the form, is executable (excluded). Three named, all sharing one property:
the demanded action ENDS the run, so nothing is presupposed of an absent reader.

• .claude/agents/os-dev.md:31 — 探针遇 403 ⇒ 停下报 blocked,⛔ 不进重试循环;只有网络错误才退避重试。
  「停下报」 reads like 「先报告」, but blocked is a named terminal report status and nothing is
  demanded after it.
• .claude/skills/pm-dispatch/SKILL.md:540 — dev 装不下预算 ⇒ 停手回报,⛔ 不自行扩写、不自行抬预算。
  Same shape, dev-addressed, no continuation.
• .claude/agents/os-dev.md:356 — 每个方案逐轴分析,推荐也按那些轴给理由;派发词没带,停下向 PM 索取,⛔ 不自拟一套轴。
  The sharpest control: it literally names an ask directed at the PM. It is still excluded, because
  「停下」 makes the ask the run's final act rather than a mid-run one. This is the discriminator the
  criterion supplies — 「再继续」, the continuation, is what makes a hit, not the presence of an ask.

Direction B — a live positive. #14031's own measured sentence (「动手前先报告」, from dispatch word
#13981) does NOT survive on these surfaces: it lived in a dispatch word, never in the skills corpus.
Probe grep -n "动手前先报告" over all 25 files returns zero, and the same grep run for the control word
「报告」 matches on 110 lines across 12 of them, so the zero is a reading and not a broken probe.

The nearest surviving relative is hit 2 / hit 3, and it is structurally identical to the retired example:
「动手前先报告」 sequenced a report before an action; 「报备后按生产者侧修」 sequences a notification
before an action. Same presupposition, same three-way collapse. The reading is live.

Coverage — so silence can be told from absence

Instruction unit = one bullet or numbered line. Re-derivable over the 25 files with
grep -cE '^[[:space:]]*(-|[0-9]+\.)[[:space:]]' FILE.

population	count
instruction units examined, all 25 files	2283
of those, addressed to a one-shot executor (the population the criterion can apply to)	440
— .claude/agents/os-dev.md, whole file	327
— skills/objectstack-pm-dispatch/rules/dev-template.md, whole file	17
— internal SKILL.md + references, lines naming dev as the actor	94
— published SKILL.md, the two dispatch-prompt non-negotiables	2
hits	3
examined within that population and judged NOT of the form	437
excluded: addressee is a session-type seat (PM / triage / director) that does have a mid-run channel	1843

Plus one prose block examined outside the bullet count: the published SKILL.md decision-frame stop
trigger (lines 630-636), which the PM pastes into every dev prompt. Judged not of the form — it returns
needs_decision, which is terminal.

The 94-line dev-addressed subset inside the PM corpus is a mechanical proxy (lines naming dev), read
individually; the addressee judgement itself is not mechanical, and the exclusion of 1843 seat-addressed
units rests on it. Stated so a reader can retest it rather than inherit it.

The exact file list counted

Worktree objectstack-issue-14768, detached at main = cc5b3dd0c2, with the thirteen frozen files
checked out from PR #15641 head 4008475c8 (marked ✦).

file	blob	lines
.claude/skills/pm-dispatch/SKILL.md	520c35c49a0aabcd0da6b04368ba298ac1662f8e	811
.claude/skills/pm-dispatch/references/compile-surfaces.md ✦	cf33306335798763c0be6bbf90280eb4b717001f	24
.claude/skills/pm-dispatch/references/contract-review.md	615acdbc13ba3395dd09ba5b3a5e8d3033b43866	60
.claude/skills/pm-dispatch/references/core-rules.md	e53c815271a412187ada33390a2f6ef349904d01	150
.claude/skills/pm-dispatch/references/decision-analysis.md	54a0bcbdabceb28656427d11b54cb2122bacf4e2	50
.claude/skills/pm-dispatch/references/dispatch-runbook.md	f8d8d730bc0cc3829684bc34fcfc1127ebe0a8de	241
.claude/skills/pm-dispatch/references/landing-operations.md	49b779ca6d106c11a7cc2476563e6a3f0bf7e75b	69
.claude/skills/pm-dispatch/references/lanes/cli.md ✦	c624e5ecce86261286bca7bcee530748a18b5460	29
.claude/skills/pm-dispatch/references/lanes/devx.md ✦	3d5a319335e1fe817e10583886b1300983a76334	38
.claude/skills/pm-dispatch/references/lanes/director.md ✦	5d88067df31979e4ad0604240604f9661032b77e	72
.claude/skills/pm-dispatch/references/lanes/engine.md ✦	a35d891f70e3519029da9f7bc39f19c7b93c87c2	32
.claude/skills/pm-dispatch/references/lanes/hotcrm.md ✦	b42e008e35f733a632816ef274e645794d6343d7	53
.claude/skills/pm-dispatch/references/lanes/services.md ✦	80c3edbe06f7175edf2ea31bc30195a046249a23	27
.claude/skills/pm-dispatch/references/lanes/skills.md ✦	2daeccf0f65e8973248f4e8b93aebac835977574	33
.claude/skills/pm-dispatch/references/lanes/spec.md ✦	a4701cd9799b76234813ccf5256133f0f7ab59bd	43
.claude/skills/pm-dispatch/references/platform-readings.md ✦	404ed68944e648590104de39ca1d6dd7eb94cd76	359
.claude/skills/pm-dispatch/references/release-aftercare.md ✦	a3a0b4590d0dd5c975099bd32b633cb2d64f024b	50
.claude/skills/pm-dispatch/references/rest-channel.md ✦	64af55ce302dad51466158658fbb4d6d7bdc8b58	82
.claude/skills/pm-dispatch/references/review-checklist.md	ffbe422d8a459ef4caa5fadaa768783b19296227	77
.claude/skills/pm-dispatch/references/seat-post-protocol.md	833db0091c20e0747ca1e0782ad428c5703fd00c	91
.claude/skills/pm-dispatch/references/state-machine.md	32063ad7bd220334eeca1321689dd3eeec1d5bbb	42
.claude/skills/pm-dispatch/references/true-green.md ✦	465ffdfe3b7c6378b304a1f2d533528864099af0	32
.claude/agents/os-dev.md	b2913e631a89cb3cfbacc6de5ebaeed04dc86d6a	403
skills/objectstack-pm-dispatch/SKILL.md	3cbbed0f501736a0db3a5ecdd9e6cda43d44b434	726
skills/objectstack-pm-dispatch/rules/dev-template.md	c3be6120fbb77f5b314081799a27c36c589a9f03	130

25 files. skills/objectstack-pm-dispatch/ exists on main and holds exactly two files (entry plus
rules/), as the card's surface list predicted; nothing was invented and nothing counted zero by default.

What the count says about the back-catalogue

Three hits in 440 dev-addressed instruction units, and two of them are one clause with two homes.
The back-catalogue this card was opened to size is small — smaller than the card's own warning allowed
for. The reason is visible in the corpus: the criterion has already been applied ahead of the census.
dispatch-runbook.md:120 (⛔ 不为提问或中期汇报结束回合;开放选择按裁决与评估轴自裁,记入终报 open_questions。)
and dispatch-runbook.md:196 (⛔ 不写成发现消费者就停下:那把判决权交回被停的一方。) are the
anti-form written as standing rules, and the rules-only rewrites of SKILL.md, the seven core references
and os-dev.md carried the criterion into each member's brief, as this card's own 2026-09-04 disposition
recorded. The census measures the residue, not the original population.

Report JSON

{
  "issue": 14768,
  "status": "done",
  "branch": "none",
  "pr": null,
  "premise_still_valid": true,
  "summary": "Census half only, read-only: no file written, no PR. Counted 25 files across the four surfaces the dispatch names (pm-dispatch SKILL.md, references including lanes, .claude/agents/os-dev.md, published skills/objectstack-pm-dispatch), at main cc5b3dd0c2 with the thirteen frozen files from PR 15641 head 4008475c8. Applied the landed PR-14697 criterion (dispatch-runbook.md:197) without redefining it and found 3 hits: os-dev.md:38 (forbid) and one clause with two homes, dispatch-runbook.md:212 plus SKILL.md:542 (allow-with-trace). None of the three sits in a frozen file, so the reading does not depend on PR 15641 landing. Two scope notes for the seat: the dispatch narrowed the card's surface list by dropping the four small internal skills under .claude/skills/, which are therefore uncensused; and the low hit count is explained inside the corpus, since the anti-form is already written as a standing rule at dispatch-runbook.md:120 and :196 and the rules-only rewrites carried the criterion ahead of this census.",
  "tests": "No build or test applies to a read-only census; no verification lock taken. Evidence is the reading itself. (1) Corpus fixed and re-derivable: worktree detached at origin/main cc5b3dd0c2, then git checkout 4008475c8 -- for the thirteen frozen files; git status showed exactly 13 modified files, matching the dispatch's count. Every counted file's blob hash from git hash-object is in the comment table. (2) Coverage denominator mechanical: grep -cE '^[[:space:]]*(-|[0-9]+\\.)[[:space:]]' over the 25 files totals 2283 instruction units; the one-shot-executor subset is 440 (os-dev.md 327 + dev-template.md 17 + 94 lines naming dev as actor in the PM corpus + 2 published dispatch-prompt bullets); 3 hits, 437 examined-and-excluded within that subset, 1843 excluded as seat-addressed. (3) Seed sweeps run and every match read in context, not counted blind: 先+(报告|确认|问|请示|停|答|验|核|读|试|探), 等待/等读数/回报后/报备/请示/询问/发问/去问/索取, 回报/汇报/上报/告知/确认后/答复后/批准/授权后, 停手/停下/停止, and English report-before / confirm-before / ask-first / wait-for / do-not-proceed / round-trip / permission. (4) Positive control for direction B: grep -n for the #14031 measured sentence returns zero over all 25 files, while the control word 报告 matches on 110 lines across 12 of them, so the zero is a reading and not a dead probe. (5) Negative controls, direction A: os-dev.md:31, SKILL.md:540 and os-dev.md:356 all carry an ask-or-report shape yet end the run, and os-dev.md:356 names an ask at the PM explicitly, which is the sharpest exclusion.",
  "mcp_calls": "0 — every GitHub read went through unauthenticated repo-scoped REST (probe returned 200, /rate_limit core 15000 confirming the proxied installation credential); the report comment is one REST POST.",
  "open_questions": [],
  "out_of_scope_findings": [
    "not filed (dispatch forbids new cards this flight) — the card's own surface list names 'the four small internal skills under .claude/skills/' (checklist-author, checklist-test, dogfood-verification, spec-property-retirement) but the dispatch ruling narrowed the surfaces to four and dropped them; those four are therefore uncensused, and the seat should decide whether the census is complete without them",
    "not filed (low confidence, needs the seat's dedup) — references/decision-analysis.md:13 says 呈完即停等回批 while SKILL.md:38 says ⛔ 永不整席等维护者答复; these reconcile if the former governs batch pacing rather than seat blocking, which the same line's second half supports, so this is an ambiguity in seat-facing text and not a defect in dev-facing text — out of this census's population either way"
  ]
}

---

Generated by Claude Code

[claude]

os-dev-report addendum — the four internal skills

Extends the census posted earlier on this card to the surface the seat's brief wrongly dropped: the four
small internal skills under .claude/skills/. Same commit cc5b3dd0c2, same criterion, same columns,
same denominator method, same two-direction controls. Read-only: no file written, no PR, no new cards.

Scope check first. The four skills carry no references/ directory — each is a single SKILL.md.
find .claude/skills/{checklist-author,checklist-test,dogfood-verification,spec-property-retirement} -type f
returns exactly 4 files, so the addendum surface is 4 files and not a tree.

file	blob	lines
.claude/skills/checklist-author/SKILL.md	e488841d6c1b5a1397702ca43ec0f4be10743672	53
.claude/skills/checklist-test/SKILL.md	af42edace8044398a87477c3bd06f7ae5f1e8fc6	234
.claude/skills/dogfood-verification/SKILL.md	5083aa42e3196f5d59be4f35626da7f56ff55d79	148
.claude/skills/spec-property-retirement/SKILL.md	c8203146783c016ee8bb07c6ab9232b9cb5a80db	337

The measured list — 1 hit

#	file:line	blob	sentence (verbatim)	form	reason
4	.claude/skills/checklist-test/SKILL.md:228-230	af42edac	**认证 / 授权漏洞的复现,永不发布到 GitHub 的任何地方** —— 不进运行 issue,不进跟踪卡,不进评论。记下项 id、条款与 detail withheld pending maintainer,复现留在会话里,停下等维护者。	forbid	「停下等维护者」 demands a wait, which is the presupposed mid-run continuation itself, not merely an ask; and this skill's own §2 hands the text to one-shot runner subagents by name (派发 runner 子代理时…每个给:该项 JSON、RUNNER.md、dogfood 技能、自己的端口/DB、结果不进仓规则), so the addressee is exactly the executor the criterion names. 「复现留在会话里」 compounds it: a one-shot runner's session ends with the run, so the repro is not withheld but destroyed.

Hit numbering continues the earlier comment (hits 1-3 there, hit 4 here). Cumulative across both halves
of the census: 4 hits — 2 allow-with-trace, 2 forbid.

Hits per file: checklist-test/SKILL.md 1 · the other three files 0.
Hits per form (this addendum): forbid 1 · allow-with-trace 0.

⚠️ Operational note, not a proposal: this clause's own text says it is
同一条规则写在两处 with RUNNER 规则 2, and docs/qa/platform-checklist/RUNNER.md is not a census
surface. Whoever executes the rewrite has to move the twin in lockstep or the two copies diverge — the
census cannot see the other copy, and says so rather than implying the site is single.

Controls, both directions

Direction A — looks like the form, is executable (excluded). Three named. All three turn on the same
discriminator: the party that answers is inside the run.

• .claude/skills/dogfood-verification/SKILL.md:100-102 —
  **断言任何「缺失 / 坏了 / 不可达」的发现之前**,用**截图**(视觉真相)或**权威服务端响应**(/api/v1/meta/...、/api/settings —— 服务器真正下发的元数据)确认。永不凭一次导航后的 DOM dump 上报严重发现。
  This is the 「先确认,再继续」 wording verbatim, and it is still executable, because the confirming
  party is the executor itself: it takes the screenshot, it calls the endpoint. The criterion is about
  who answers, not about the presence of a "confirm first".
• .claude/skills/spec-property-retirement/SKILL.md:66 — 所以:**先尝试删除,让构建来裁。**
  The sharpest of the three: it explicitly names an adjudicator, and the adjudicator is a tool the
  executor runs.
• .claude/skills/checklist-author/SKILL.md:49-50 —
  **没有子代理工具时,按角度顺序跑 —— 并在交付里声明这一轮是顺序执行的。** 会话里不存在 Task/子代理工具是允许的退化路径,不是阻塞
  A degradation that must be declared. The declaration lands in the delivery, so it is terminal, and the
  run degrades under its own authority instead of waiting for permission to degrade — which is what the
  form would have produced here.

Direction B — a live positive, and a discriminating one. #14031's own measured sentence is again
absent: grep -n "动手前先报告" over the four files returns zero, while the control word 「先」 matches on
36 lines across all four, so the zero is a reading.

The re-derived positive on these surfaces is hit 4 itself, and it is a stronger relative than the one the
first half found, because it carries the wait verbatim (「停下等维护者」) rather than only a
report-then-continue sequencing. It also proves the reading discriminates rather than keyword-matches:

sentence	prefix	verdict
checklist-test:228-230 — 复现留在会话里,停下等维护者。	停下	hit — the demanded act is waiting
os-dev.md:31 — 探针遇 403 ⇒ 停下报 blocked	停下	excluded — the demanded act ends the run

Same two-character prefix, opposite verdicts, decided by the continuation and nothing else. That is the
criterion doing work, not a grep.

Coverage — so silence can be told from absence

These four are prose skills, not the rules-only one-rule-per-line corpus the first half counted, so the
denominator method transfers only partly and the honest form is two counters, stated separately rather
than blended: bullet or numbered units (mechanical, same regex as the first half,
grep -cE '^[[:space:]]*(-|[0-9]+\.)[[:space:]]' FILE, fenced blocks excluded), plus directive prose
paragraphs (a non-bullet, non-heading, non-table block carrying an imperative or prohibition marker).

file	bullet units	directive paragraphs	units
checklist-author/SKILL.md	3	4	7
checklist-test/SKILL.md	26	18	44
dogfood-verification/SKILL.md	28	7	35
spec-property-retirement/SKILL.md	41	13	54
total examined	98	42	140

population	count
instruction units examined, all 4 files	140
of those, addressed to a one-shot executor	120
— dogfood-verification, whole file (os-dev is pointed at it by name for browser verification)	35
— spec-property-retirement, whole file (a dev playbook for implementing a retirement card)	54
— checklist-test sections 2, 3, 4 and Guardrails (the text §2 hands to each runner subagent)	28
— checklist-author, the hunter/writer constraints (lines 39-42)	3
hits	1
examined within that population and judged NOT of the form	119
excluded: addressee is the orchestrating session (selector resolution, run planning, closure duty, sweep scale)	20

One borderline, recorded rather than silently decided. checklist-test/SKILL.md:152 —
同一份逐条款表 + 环境准备与测试的耗时之比,回报给维护者(chat),并链接已立的 issue。 — presupposes a
chat channel a one-shot runner does not have. It is excluded because the demanded act is the run's final
delivery and nothing is demanded after it: this is a channel mismatch, not a continuation mismatch, and
the criterion governs the continuation. Its addressee also reads more naturally as the orchestrator, which
does hold that channel. Both readings are stated so the boundary is visible rather than inferred.

Seed sweeps run over the four files

先+(报告|确认|问|请示|停|答|验|核|读|试|探) · 等待/等维护者/等人/等回/回报后/报备/请示/询问/发问/去问/索取/停下/停手/停止/待维护者/听候 ·
之前/前先/再往下/再继续/然后再 · English report-before / confirm-before / ask-first / wait-for /
do-not-proceed / pending / withheld / permission. Every match read in context. The wait-and-ask family
returns exactly one site across all four files — checklist-test:230 — and every 「先X」 match has X
self-serviced by the executor (read a doc, take a screenshot, query an API, run a build, check a port,
try the deletion and let the build rule).

What the addendum changes about the first half's conclusion

Nothing about the criterion or the totals method; one thing about the shape of the residue. The first
half's three hits were all report-then-continue sequencing, the mildest form. This one is the strong
form — an explicit wait — and it sits in the one place where obeying it literally is worst: a security
finding, where the alternative to waiting is a repro that dies with the session. The back-catalogue is
still small (4 hits over 560 one-shot-executor instruction units across both halves), but it is not
uniformly mild, and the rewrites should not be sequenced as if it were.

---

Generated by Claude Code