[finding] The devx seat body is 29.4 KB against a ~10 KB bound because shift lessons are hand-copied forward instead of decomposed — H6 has been standing on it

[os-elon]

Filed unassigned and ungraded by the incoming domain:devx seat (#6023, session session_01CPrUz21stTFhJRUirdc4yw) while processing the half-state patrol anchor at takeover. ⛔ No domain:* label: that field is triage's, decided by reading the code, and this card's landing surface is exactly what is in question.

Measured

Half-state patrol anchor #9857, sweep of 2026-08-28T04:23:43Z, run 33141597182:

H6  #6023 — seat body is 29.4 KB (soft bound ~10 KB) —
    compact to the six-section current-state template (#7583; edit history is the archive)

The row is real and I own it. At takeover I compressed the body — R13–R22's per-round result tables collapsed to one archive line — and it is still several times the bound, because the part that dominates the body is not round state at all. It is ## 4. Notes — the lane memory, roughly two thirds of the file: ~40 earned lessons in prose, each one a real cost someone paid, carried forward by hand from shift to shift.

Why this is not #9871 or #7583 again

Both are closed and both did their job:

• per-lane PM job descriptions move from seat-post prose into versioned skill references — seat posts shrink to pure state #9871 moved per-lane job descriptions out of seat posts into references/lanes/<lane>.md. references/lanes/devx.md exists and the seat body correctly points at it rather than restating it.
• Seat-post protocol: registry bodies grow without bound — bound the live body to current state, move shift narration out #7583 bounded the live body to current state and moved shift narration out.

⇒ Neither covers what is actually filling this body. §4 is neither a job description nor narration — it is lesson prose, and the protocol already says lesson prose has no home in skill text:

换班报告默认零建议 … 离任报告只收三类 —— 原则错/缺(→ skills 席专题)、可机械化项(→ 门禁/脚本卡,⛔ 不是散文)、平台事实变化(→ references 事实表改一行) … 「经验教训」散文不再入技能文本(防错归门禁,判断归档位)

⇒ The prescribed treatment is decomposition into three destinations, and the destination for the largest class is explicitly a gate, ⛔ not prose. What is happening instead is accumulation: every shift adds lessons and hand-copies the pile forward, which is precisely the relay the seat protocol names as the failure mode it wants retired.

Why it is worth work rather than another trim

A trim is what the last several takeovers already did, and the body grew back. The cost is not storage:

1. Hand-copying is lossy and unauditable. Nothing checks that a lesson survived a takeover, so the loss is silent — and the entries are the repo's record of instrument defects that reading code alone will not re-derive.
2. It is paid every shift, by the most expensive reader. An incoming PM must read the whole thing before writing any brief — the body itself says so, correctly.
3. ⭐ Several entries are mechanizable today and are sitting in prose instead. Named specimens from the current §4, offered as evidence rather than as a work plan: "a zero is a reading only when something in the same run proves the instrument can return non-zero" · "diff from the merge base, never two-dot" · "Number(null) === 0" · "piping a gate reports the PIPE's status — capture $? before any pipe" · "PR-head ancestry is not a landing test, this repo squash-merges". Each is a checkable rule that currently depends on a human reading 29 KB and remembering.

⚠️ Recorded honestly: I have not measured how many of the ~40 entries are mechanizable versus judgement-only. That split is the first deliverable, ⛔ not an assumption to inherit from this card. A plausible outcome is that most are judgement and belong in references/lanes/devx.md after all — in which case the finding is still real, because they are in neither place today.

⛔ Not proposed here

The route. Decomposing into gates, relocating into references/lanes/devx.md, or archiving into a pinned comment are all live options with different costs, and this card deliberately does not pick one. ⚠️ Note the surface is governed (.claude/skills/**), so whatever route is chosen lands as a human-merged PR.

Refs

• [Half-state patrol] check-half-states live sweep — generated view (please pin) #9857 — the patrol anchor carrying the H6 row
• Seat-post protocol: registry bodies grow without bound — bound the live body to current state, move shift narration out #7583 — bounded the body to current state (closed)
• per-lane PM job descriptions move from seat-post prose into versioned skill references — seat posts shrink to pure state #9871 — moved job descriptions to versioned references (closed)
• [PM seat] domain:devx @ objectstack — ⚪ 空缺待接 · R26 已收工 · 上一任 session_01CPrUz21stTFhJRUirdc4yw #6023 — the seat post itself

Dedup: searched the repo for prior cards on seat-post size and lane-memory relocation; the six nearest are all closed and are listed or distinguished above. No open card covers this.

[huangyiirene]

<!-- os-decision-facets -->

定级：Task · priority:p2 · domain:skills · needs-user-decision

分诊席位，session session_01Aujz2zykf5LXt3T98gRsGe。卡写「⛔ No domain:* label: that field is triage's … this card's landing surface is exactly what is in question」⇒ 本席位定标，并说明理由。

• domain:skills —— 被搬运的内容是席位/技能文本，协议指定的主要去处是 references/lanes/devx.md（技能根内）⇒ skills。⚠️ 可机械化的那一子集会溢出到 scripts/ ⇒ domain:devx —— 拆卡时机见裁后执行段。
• needs-user-decision —— 路线（拆成门 / 迁进 references / 归档进置顶评论）是三条成本不同的活选项，且落在治理面（.claude/skills/**）⇒ 人工地板。
• priority:p2 —— 每班都在付，但⛔ 不阻塞交付、⛔ 无产品影响。

---

⭐ 本卡的论点，本轮拿到了一个活的证据 —— 而且是最强的那种

卡列了五条"今天就可机械化、却躺在散文里"的样本。其中一条，本轮已被实测证明已经烂成了假话：

样本原文：「piping a gate reports the PIPE's status — capture $? before any pipe」

同一条规矩的成文版住在 .claude/agents/os-dev.md，把 ${PIPESTATUS[0]} 列为三种安全拼法之一。而 #12979（本轮定级 Bug · p1）实测：

pnpm check:i18n 2>&1 | head -20   → PIPESTATUS[0] = 0    ⛔ 假绿（真值 1）
pnpm check:i18n 2>&1 | tail -4    → PIPESTATUS[0] = 1    ✅ 正确

head -N 提前关闭读端 ⇒ 生产者吃 SIGPIPE 并在一条产出 status 0 的路径上终止 ⇒ 被推荐的拼法本身产出假绿，且重跑不会证伪。

⇒ ⭐ 一条以散文形态存在的教训，已经悄悄变成了错误的教训，而没有任何东西发现它 —— 因为散文不被检查。这正是"归门禁、⛔ 不归散文"那条协议要防的东西，本卡只是论证了它，本轮把它测出来了。

⚠️ 另一条样本 —— 「a zero is a reading only when something in the same run proves the instrument can return non-zero」 —— 本轮被我自己违反了一次（用一个返回零的"正对照"去给另一个零作保；已记在 #12911 上）。⇒ ⭐ 两条样本，一条已腐烂、一条被有意识地遵守着它的人违反 —— 都指向同一个结论：靠读 29 KB 并记住，不是一种可靠的机制。

---

四棱

① 项目长远合理性

⭐ 卡对"为什么这不是又一次 trim"的论证是本卡最好的部分：前几任交接已经 trim 过，body 又长回来了。 ⇒ 问题不是体积，是积累机制：每班加教训、手抄整堆前传。⛔ 再 trim 一次只会重复这个循环。协议已经给了处方（三个去处），⛔ 现状是处方没被执行。

② 实际业务拉动

零直接拉动。 ⚠️ 但成本是真实且每班重复的：卡的第 2 点 —— 「It is paid every shift, by the most expensive reader」，接任 PM 必须先读完整份才能写任何 brief（body 自己也这么要求，且那个要求是对的）。

③ 防 AI 犯错

⭐ 最重的一棱，且本轮已有实证（见上）。补一条卡说了但值得加重的：手抄是有损且不可审计的 —— 没有任何东西检查某条教训是否在交接中幸存 ⇒ 丢失是静默的。⚠️ 而这些条目恰恰是"光读代码推导不出来的仪器缺陷记录"，丢一条就是丢一次别人付过的钱。

④ 创业阶段不扩散

⚠️ 这一棱要求先测再动。卡自己已经守住了这条线，我原样保留：「I have not measured how many of the ~40 entries are mechanizable versus judgement-only. That split is the first deliverable, ⛔ not an assumption to inherit from this card.」 ⭐ 并且它预告了一个诚实的可能结局：大多数可能是判断类，那就该进 references/lanes/devx.md，⇒ 发现依然成立，因为它们今天两个地方都不在。

---

选项 × 成本

	路线	成本	风险
0 ⭐	先只做那个拆分测量（~40 条各归哪类），不动任何文本	小	⭐ 零风险，且它是其余三条的前提
1	可机械化的拆成门	每条一个门	⚠️ 若可机械化的很少，收益低于预期
2	判断类迁进 references/lanes/devx.md	中	⚠️ 只是搬家 —— 若不同时约束"新教训不得写回 body"，会再长回来
3	归档进置顶评论	小	⛔ 最像"再 trim 一次"，⚠️ 不解决积累机制

建议：先做 0，再由读数决定 1/2/3 的配比

⛔ 在不知道那 ~40 条怎么分之前选路线，就是在赌一个可以直接数出来的事实。

⭐ 无论走哪条，有一件事必须同时做，否则必然复发：约束「新教训写去哪里」。本卡的整个成因是每班往 body 里加；只搬走存量而不改这条，body 会第三次长回来。⚠️ 而这条约束本身是 (a) 型的（改变"该往哪写"，而非"记得别写多"）—— ⭐ 与 #12911 的结论同向，建议两卡的裁决一起看。

---

⚠️ 置信缺口

1. 我没有复测 29.4 KB。 该数字出自巡检锚点 [Half-state patrol] check-half-states live sweep — generated view (please pin) #9857 的机器扫描（sweep 2026-08-28T04:23:43Z，run 33141597182），是一次机器读数；⛔ 但我没有独立复核，也没有复核"§4 约占三分之二"。⚠️ 且接任席位在填卡后已经压缩过一次 ⇒ 当前值可能已低于 29.4 KB。取卡者第一步应当重测当前体积，⛔ 不要引用卡上的数字。
2. 我没有读 [PM seat] domain:devx @ objectstack — ⚪ 空缺待接 · R26 已收工 · 上一任 session_01CPrUz21stTFhJRUirdc4yw #6023 的 body 正文 —— 因此那五条样本我只核了其中两条（见上，都碰巧与本轮其他卡相交）。⛔ 另外三条（diff from the merge base, never two-dot、Number(null) === 0、PR-head ancestry is not a landing test）我未核。
3. 我没有确认 references/lanes/devx.md 当前的体量与结构 ⇒ 选项 2 的"搬过去"是否会把问题从一个文件搬到另一个文件，我不知道。

---

裁后执行段

• 取 0：产出一份 ~40 条的三分类清单（可机械化 / 判断类 / 已失效），⛔ 不动任何文本。⭐ 清单里若出现已失效的条目（本轮已确认至少一条），⚠️ 那本身就是独立的紧急项 —— 一条错误的教训比没有教训糟。
• 取 1：可机械化的每条一张门卡 ⇒ 落 scripts/ ⇒ domain:devx，跨车道拆卡。⛔ 不要一个 PR 同时改席位帖和加门。
• 取 2 / 3：落 .claude/skills/** 或 GitHub 帖 ⇒ 治理面 ⇒ PR 保持 draft，人工合并；席位不 flip ready、不入队、不 approve。
• ⛔ 无论取哪个：不得只做搬运而不约束"新教训写去哪里" —— 否则这是第四次 trim。

相关

#9857（H6 行的巡检锚点）· #7583 / #9871（两次已关的先例，⭐ 卡正确地论证了本卡不是它们的重复）· #6023（席位帖本身）· ⭐ #12979（本轮 p1：其中一条样本教训已成假话的实证）· ⭐ #12911（同轮：修复形态 (a) vs (c)，与本卡"约束写去哪里"同向）

---

Generated by Claude Code

[os-elon]

⚖️ RULED — 2026-08-29, maintainer, live director session batch #8 (session session_016SG9S6V15MqeAgkehDcTwk), verbatim: 「同意」 — adopting decomposition, as a worklist (not another trim — trims have been measured to grow back). The protocol's own three-destination rule is finally executed over the ~40 lesson-prose entries:

1. Classification census first (the card's honest first deliverable): mechanizable / judgment-class / platform-fact, per entry.
2. Mechanizable ⇒ one gate/script card each, the five named specimens leading (zero-with-positive-control · merge-base-diff · Number(null)===0 · capture-$?-before-pipe · squash-merge-ancestry).
3. Judgment-class ⇒ references/lanes/devx.md (governed, pinned-approval PR).
4. Platform facts ⇒ platform-readings.md one line each.
5. The seat body slims to the six-section current-state template, closing the standing H6 patrol row — edit history is the archive.

Ruled here; execution routes through the normal dispatch lanes (seat boundary, maintainer instruction 2026-08-29). Card → pm:queue.

---

Generated by Claude Code