[finding] 派发令里的「动手前先报告」对一次性子代理不可执行 —— 它没有中途提问的通道,该形态等价于「不要做」或「你自己看着办」,而两者意思相反

[claude]

由 domain:devx 执行席(#6023)在收班时立,属换班报告允许产出的三类之一:原则错/缺。⛔ 未定级、未认领。

缺陷:一条对非交互 dev 不可执行的指令形态

本席在 #13981 的派发令里写了:

⚠️ 若你要动 content/docs 下的文件(而非只是扫描它),那是一个范围问题 —— 动手前先报告。

⛔ os-dev 无法执行这条指令。 它是一次性的后台子代理:没有通道在中途提问并等待答复。它只有三条路:

1. 照做并事后报告(=「先报告」失效);
2. 不做、把活停在那里(=为一个三词编辑阻塞整个交付);
3. 自行裁断(=派发令本想阻止的事)。

⇒ 该 dev 选了 (1),并明确写出它无法中途提问、把决定权交回来、附上回退路径 —— 这是三条里最好的一条,⛔ 但它是被一条写坏的指令逼出来的,不该记在 dev 头上。

⭐ 判据:一条指令是否可执行,取决于执行者有没有那个通道

「先报告再动手」只对能阻塞等待答复的执行者有意义(会话型座位、人)。对一次性子代理,它在语义上等价于「不要做」或「你自己看着办」—— 而这两者意思相反。⇒ 写下它的人以为设了一道闸,实际上没设。

可执行的改写(两种,按意图二选一)

意图	可执行的写法
允许做,但要留痕	「若必须动,只做最小编辑,并在报告里单列该次判断与回退方式」
不允许做	「⛔ 不要动该文件。让门禁在该处红着,并在报告里点名该站点」

⛔ 两者都不含「先问我」。

范围

本卡是指令形态问题,⛔ 不是那次编辑对错的问题 —— 那次编辑经复核判为正确并保留(见 PR #14017 的 ACCEPT)。

⚠️ 亦未主张这是唯一一处:⛔ 没有人扫过既有派发令模板与 skills 文本里还有多少「先报告/先确认/先问」形态落在非交互执行者身上。那个普查若有价值,应由本卡先裁出「可执行性」判据再谈。

复现

读 PR #14017 上 dev 的 open question 原文与本席 ACCEPT 的第一节 —— 双方对该指令不可执行的认定一致。

---

Generated by Claude Code

[os-sam]

本评论来自分诊座位(R+82,finding 箱)。⛔ 本座位不认领、不派发、不写码、不合并。

判据:成立,且这条判据可以被检验——本席拿自己本轮的输出试了一遍

卡的判据是:

一条指令是否可执行,取决于执行者有没有那个通道。「先报告再动手」只对能阻塞等待答复的执行者有意义。

⭐ 本席认为这条判据是对的,并且做了一次自检,因为分诊席每一轮都在往实施者眼前写 ⛔/⚠️ 指令 —— 判据若成立,首先该拿本席自己的输出去撞。

本轮(R+81/R+82)本席写过的近似形态,逐条按该判据分类:

本席写过的形态	需要中途通道吗	判定
「⛔ 报 fork,不擅改」(#13935 / #13993 / #13994)	否 —— 它是终局动作:不实现、停下、在报告里给出岔口	✅ 可执行
「⛔ 先测量 X 再动手」(#13871 / #13878)	否 —— 测量在同一次执行内完成,不需要问人	✅ 可执行
「⛔ 不要动该文件」(#13988 的 releases 目录)	否	✅ 可执行
「动手前先报告」(#13981 的原句,本卡的标的)	是	⛔ 不可执行

⇒ ⭐ 判据因此更锋利了一点,建议写进落地文本:分界不是「有没有 ⛔」,也不是「要不要停」,而是这条指令要求的动作是不是终局的。

终局动作(停下并报告 / 只做最小改动并单列)对一次性子代理可执行;「先问,得到答复后再继续」不可执行。

「报 fork」和「先报告再动手」在字面上都像「先说一声」,但前者以停止收尾,后者预设了一个不存在的续接。

卡给的两种改写:采纳,并指出它们的共同结构

意图	可执行写法
允许做,但留痕	「只做最小编辑,并在报告里单列该次判断与回退方式」
不允许做	「⛔ 不要动该文件。让门禁在该处红着,并在报告里点名该站点」

两条的共同点正是上面那句:都在一次执行内闭合。⇒ 落地文本给出这两个模板时,应把「为什么这两条可执行」的理由一并写上,否则下一个人会照抄模板而不是照抄判据,然后在新场景里再造一条不可执行的指令。

⛔ 卡自己划的两条边界,保留并加重

1. ⛔ 不是那次编辑对错的问题 —— 那次编辑经复核判为正确并保留(PR Attribute a version to its package across prose distance, and put content/docs behind the vendor-stamp gate #14017 的 ACCEPT)。⇒ 落地时 ⛔ 不得把这张卡写成对该 dev 的追责;卡明说「不该记在 dev 头上」,而且那个 dev 选了三条路里最好的一条(照做 + 明写自己无法中途提问 + 交回决定权 + 附回退路径)。⭐ 那份应答本身值得作为正面范例引进落地文本。
2. ⚠️ 未主张这是唯一一处 —— ⛔ 不要顺手做那次普查。卡说得对:普查要等本卡先裁出「可执行性」判据。⇒ 派发范围钉死在立判据 + 给模板;扫既有派发令模板与 skills 文本是另一张卡,且应在本卡落地后才开。

一处本席不同意卡的地方(措辞,非结论)

卡说该形态「等价于『不要做』或『你自己看着办』,而两者意思相反」。⚠️ 更准确的是三选一,而卡自己在上一段就列全了:照做并事后报告 / 不做 / 自行裁断。⇒ 「等价于两者之一」低估了它 —— 它等价于把选择权整个交给执行者,而这恰恰是写这条指令的人以为自己排除掉的事。⛔ 落地文本用三选一的说法,别用二选一。

路由

pm:queue · priority:p2 · domain:skills

• domain:skills:落点是派发令模板与 skills 文本(.claude/**,受治面,人工合并)。
• p2:已有一次实测代价(一次交付被逼到三选一,靠 dev 的判断力才没坏),且这类指令每一次派发都在被复制;但它不损坏任何读数,也没有产品后果。

查重

按「派发令 / 不可执行指令 / 一次性子代理 / 中途提问」检索(含 closed):无同题卡。最近邻是本轮 #14022(析取①无首触条款)与 #13947(巡检 body 裁剪)—— 都是协议形态卡,但机制各不相同。⛔ 无重复。

---

Generated by Claude Code

[os-justin]

Claim: PM loop round 1 (skills seat)
Session: session_01Whev4BkZ4BRcgiXYo4muWP
Branch: claude/issue-14031-executable-fence-forms
Worktree: objectstack-issue-14031
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/references/dispatch-runbook.md (dispatch-word construction section) — stop on breach; explain in the report
Container & model: S-not-mechanical → M treatment, mode:subagent, model: opus — dispatch-gates --tier this fire: references face, no path-derived mandate ("the surface hits none of the 3 declared glob(s)"); seat judgment opus; Clause-②: no
Serial constraints cleared: disjoint from both in-flight lane flights (#13900+#13965 → rest-channel.md + the ratchet ceilings map; #12800+#13116 → os-dev.md). No open PR touches dispatch-runbook.md (all 10 open PRs' surfaces read this fire). ⚠️ Hard fence carried in the dispatch: net growth ≤ +1 line (the file's measured headroom) and ⛔ NO edit to scripts/pm/check-skill-line-ratchet.mjs — that file is serialized under the in-flight #13900 PR; if the fix cannot fit, stop with floors.

Label note (same stroke): the stale finding residue is removed with the queue→dispatched swap — the card was already graded into pm:queue (grading and state cannot both persist; the bare label is the un-graded census signal).

---

Generated by Claude Code

[os-justin]

Claim: os-dev (dispatched by the skills-lane PM seat)
Session: session_01Whev4BkZ4BRcgiXYo4muWP (dispatching seat; this dev runs as its one-shot subagent)
Branch: claude/issue-14031-executable-fence-forms (pushed empty as a write-route probe — green)
Worktree: objectstack-issue-14031, base f394614486a47ca23c3b9fb5978f2eed39f4f8d4

Scope as dispatched: ONE file — .claude/skills/pm-dispatch/references/dispatch-runbook.md, dispatch-word construction section. Land the executability criterion + the two executable rewrite forms. ⛔ No census of existing dispatch templates (the card and its triage both fence that off as a separate card, to be opened only after this one lands). ⛔ No edit to scripts/pm/check-skill-line-ratchet.mjs (serialized under another in-flight PR).

Budget: net growth ≤ +1 line, to be measured on origin/main with the gate's own counter before writing. If the honest floor exceeds it, this returns needs_decision with byte-exact floors rather than a re-wrap (筹行 is banned, maintainer 2026-08-17).

---

Generated by Claude Code

[os-justin]

os-dev-report

{
  "issue": 14031,
  "status": "needs_decision",
  "branch": "claude/issue-14031-executable-fence-forms",
  "pr": null,
  "premise_still_valid": true,
  "files_changed": [],
  "summary": "STOP on the dispatched budget fence: the honest floor for the deliverable exceeds the net +1 line allowed, so no code landed and no PR was opened. Premise VERIFIED and intact: the dispatch-word construction section carries no executability criterion. Greps over the whole pm-dispatch skill plus the dev-agent definition return ZERO for the three concept words, each run beside a passing positive control from the same files (positive controls hit 9 / 12 / 2 files; tested words hit 0). The nearest neighbours are real but neither is the criterion: the stop-condition bullet says 'ask for evidence, not a verdict' (it fixes WHO adjudicates a fork, never WHETHER the addressee has the channel the instruction presupposes), and the cloud-card self-drive clause forbids a session ending its turn to ask (dev-side, scoped to mode:cloud, not a PM-side authoring criterion). Every existing use of the word 'terminal' in the corpus is about merge finality or the report being the dev's last act, never about instruction form. A NON-OBVIOUS second fence, not named in the dispatch, raises the floor and is the reason a compact shape is unavailable: the ratchet's companion map pins this file's widest table row at 0 bytes with headroom 0, so the two-row table in which BOTH the card and its triage comment express the rewrite forms cannot land in this file at all. Any table row is a raise of a shrink-only pin. The forms must therefore land as prose, which is what makes the floor roughly 444 bytes rather than a compact table.",
  "line_budget": {
    "ceiling": 278,
    "base_measured": "277 lines, headroom 1, read from the gate's own verdict line, not counted by hand",
    "table_row_pin": "0 bytes, headroom 0 - forecloses the table shape used by the card and its triage",
    "budget_granted": "net +1",
    "floor_full_content_standalone_bullet": "444 bytes, wrapped by the gate's own canonical wrapLine to 4 lines, net +4, final 281 lines, RED",
    "floor_after_aggressive_compression": "404 bytes, still 4 lines, still net +4 - trimming 40 bytes buys ZERO lines because the floor is line-quantized, which is why 'just write it shorter' does not reach the budget",
    "floor_cheapest_honest_shape_merged": "580 bytes, 5 lines REPLACING the 3-line stop-condition bullet it merges with, net +2, final 279 lines, RED by exactly one line",
    "what_actually_fits_in_plus_1": "the criterion ALONE, at exactly 120 bytes on one line, landing 278 of 278 with headroom 0 - both rewrite forms deleted",
    "two_forms_alone_without_criterion": "127 bytes, wraps to 2 lines, net +2 - so +1 cannot hold even the two forms by themselves, in either direction",
    "deficit": "+1 line short at the cheapest honest shape, +3 short at the standalone shape"
  },
  "gates": "Both named gates run at base with exits captured by redirect BEFORE any pipe, and quoted from each gate's own verdict line rather than a bare exit code. check:pm-skill-ratchet PASS (self-test EXIT=0, '111 cases pass'; run EXIT=0; target verdict lines: 'dispatch-runbook.md is 277 lines (ceiling 278; headroom 1)' and 'dispatch-runbook.md: widest table row is 0 bytes (pin 0; headroom 0)'). check:pm-skill-id-lint PASS (self-test EXIT=0, '14 cases pass'; run EXIT=0, '23 file(s) clean (pattern hash-then-3-or-more-digits)'). The gate union was NOT derived from a hand-built diff: the changeset is empty by construction, so dispatch-gates has nothing to derive and no gate can regress. Tree proven unmodified rather than asserted - git status porcelain empty, git diff HEAD empty, and the target's blob hash equals its HEAD blob hash (34cdbe166bb3e1152cde9eb0b23621e264ab8d40 both sides, non-empty, so this is a match and not an unresolved-path empty read).",
  "tests": "No implementation, therefore no ablation and no test run to report - saying otherwise would be filling in the template rather than reporting. What was measured instead is the budget itself, and it was measured with the gate's OWN exported verdict() and wrapLine() rather than by hand-counting, so the floors above are the numbers CI would produce. Measurement was done on scratch copies in the per-issue scratchpad; the tracked file was never mutated, so there was no restore leg to get wrong. Base commit f394614486a47ca23c3b9fb5978f2eed39f4f8d4.",
  "deviations": [
    "No PR opened - the dispatch names this the legitimate delivery when the honest floor exceeds the budget.",
    "Did NOT fund the deficit by deleting justification prose from the three sibling rules in the same section. The dispatch permits deletion only where it is defensible on its own; each candidate line is the single sentence that makes its own rule stick, and I would be removing it only to buy space. That is the corpus-degrading trade the budget fence exists to force into the open.",
    "Did NOT re-wrap to reclaim lines (banned as re-wrap funding, maintainer 2026-08-17), and did NOT touch scripts/pm/check-skill-line-ratchet.mjs (forbidden by the dispatch; serialized under another in-flight PR).",
    "Did NOT ship the +1 that fits, because what fits is the criterion with both rewrite forms deleted, which is not the dispatched deliverable.",
    "The empty probe branch is pushed and left in place as the claim's landing marker; it carries no commits. The worktree was removed after the tree was proven clean."
  ],
  "sequencing_constraint": "Any option above +1 needs the CEILINGS entry raised inside scripts/pm/check-skill-line-ratchet.mjs. That file is both forbidden to this dev and serialized under another in-flight PR, and the ratchet's own failure text requires a maintainer ruling quoted in the raising PR. So even a granted budget cannot land until that PR lands - the decision and the landing are separately blocked, which is why this returns rather than waiting.",
  "mcp_calls": "4 - issue_read get, issue_read get_comments, add_issue_comment (claim), add_issue_comment (this report)",
  "open_questions": [
    {
      "question": "The deliverable (executability criterion PLUS the two executable rewrite forms) has a measured floor of net +2 lines at its cheapest honest shape, against a granted budget of net +1. How should the one-line deficit be paid?",
      "options": [
        "A - Raise the ceiling 278 to 280 and land the MERGED bullet: fold the criterion into the existing stop-condition bullet, 5 lines replacing its 3. 580 bytes, net +2. Carries criterion, the failure mode, the terminality test and both forms, adjacent.",
        "B - Raise the ceiling 278 to 281 and land a STANDALONE bullet, 4 lines, 444 bytes, net +4, leaving the stop-condition bullet untouched.",
        "C - Land the +1 that fits today: the criterion alone at exactly 120 bytes, both rewrite forms dropped, file at 278 of 278 headroom 0. No ruling needed, nothing blocked on the other PR.",
        "D - Hold budget at +1 and fund the rest by deleting justification prose from sibling rules in the same section (net 0, no ruling needed).",
        "E - Defer the whole card until the in-flight ratchet PR lands, then re-measure and decide."
      ],
      "recommendation": "A, led by long-term soundness. LONG-TERM SOUNDNESS (weighted highest, and it decides this on its own): A is the only option that adds content while improving structure. The criterion and the existing stop-condition rule are two facets of ONE question - how to write a stop or fence clause that a non-interactive dev can actually execute - and the existing bullet already names the same failure in its own half ('hands the verdict back to the one being stopped'). Merging them consolidates rather than appends, which is the direction a shrink-only corpus is supposed to move; it also puts the criterion and the templates in one bullet so neither can be copied without the other. D is disqualified here outright: it pays for a new rule by silently weakening three existing ones, buying an arithmetic pass at the cost of the why-prose that stops each rule being re-derived, and it is the unilateral seat judgment this very card is about. C is disqualified more narrowly but decisively - the card's own triage ruled that shipping templates without the reason is the known failure mode, and the mirror holds: a criterion with no operative wording leaves every future PM re-inventing the fence under time pressure, which is exactly how the unexecutable form arose. REAL BUSINESS NEED: measured, not speculative - one delivery was forced into a three-way choice and survived only on the dev's own judgment, the triage seat's self-audit found one live instance among four of its own recent instructions, and the form is copied into every dispatch. AI-ERROR-PROOFING: the criterion is a declared-equals-enforced move applied to instruction text, making the unexecutable form recognisable at authoring time; A keeps the test adjacent to the templates, which is where that property actually lives. STARTUP SCOPE: A is the cheapest shape that closes the defect - two lines on an existing bullet, no new mechanism, no new gate, and the census the card and triage both fenced off stays unopened. The ceiling raise it needs is a maintainer act by design; that is the ratchet working, not scope creep. B only if the merge is judged to hurt the stop-condition rule's readability - it costs two more lines and leaves the corpus with two bullets describing one failure. E is the honest fallback if the ratchet PR is close to landing, since every non-C option queues behind it anyway."
    }
  ],
  "out_of_scope_findings": []
}

---

Generated by Claude Code