Repository navigation
[Decision] Skills optimization program — batch 2 (1 item): does #14296 item 1's "≥10% off the package in the same PR" bind the split PR arithmetically, or is it the anti-growth guard? (ui split measured at −0.16%) #14568
Description
Activity
Maintainer ruling recorded — A: condition (a) of the #14296 item-1 ruling is the programme-level anti-growth guard; a split PR lands token-neutral, never positive
Director seat (objectstack #12708), summon #10, session
session_01ShyhexkB2d1AeRZ85tgAAe, 2026-09-02.Provenance (who / verbatim / where): maintainer, live PM chat with the director seat, 2026-09-02, replying to decision batch #12 in which this card was item 1 with the recommendation A (fallback C; B and D not recommended). Verbatim reply: 「同意」.
Ruled: A. The "≥10% net off the package in the same PR" condition is read at programme level: the split PR itself must be token-neutral or negative (the ui split measured −56 tokens; a positive delta is refused), and the package's arithmetic is carried by the programme's flights on that package together — for
objectstack-ui: round 1 (#14338, −2,585) + the split (−56) + the generated-artifact seam card #14559 (−5,352) = −7,993 against 36,542 before the programme, −21.9%. Each change stays in its own lane under its own owner. Options B (fold #14559's generated-artifact deletion into the ui split PR, one PR across two domains), C (wait for #14559, which collapses to A plus delay) and D (do not split) are not taken.The reading applies identically to the pm-dispatch split (#14300 round 2).
Execution: skills lane. #14298 round 2 continues on the original brief (same dev, same branch); the PR body carries the programme-level account and the −56 measurement; the ratchet's new lines are pinned at the landed count and the entry line re-locked. The three
check:skill-identifier-livenessLeg-2 path bindings stay in the entry file as headings with their bodies moved out, declared in the PR as the one-topic-two-files cost. Governedskills/**surface ⇒ draft-only, review requested fromos-zhuangandhotlong, human merge.State transition, same stroke:
needs-user-decision→pm:queue;priority:p1,domain:skillsretained. Ledger: objectstack director seat post #12708, summon #10.
Generated by Claude Code
Ruling A recorded and executed by the skills lane seat (session
session_01LraLgQVGq8egUwfYZpbYt1, 16:4xZ): the three cards this batch gated are unblocked in the same stroke — #14298 (ui split)pm:blocked→pm:dispatched, same dev resumed on the same branch with the programme-level account; #14300 (pm-dispatch split)pm:blocked→pm:dispatched, round 2 claimed; #14297 (datarules/security.md)pm:blocked→pm:queue, next free slot. The ruling text lives in the director's comment above and is quoted on each card. Closing this decision card as completed;pm:queueremoved — execution is tracked on the member cards.
Generated by Claude Code
- added a commit that references this issue
on Sep 2, 2026
Decision batch 2 of the skills catalog optimization program #14292 (maintainer mandate 2026-09-02, verbatim: 「审核所有的 skills，进行全面的优化。」). Filed by the skills lane seat (session
session_01LraLgQVGq8egUwfYZpbYt1). One item; a one-line reply (e.g.1A) executes it. Nothing else in the program waits on this: the other flights, the stub deletions and the seam card #14559 proceed.Measured by the #14298 round-2 dev at objectstack
origin/main310dd01a(report on #14298, commentos-dev-report, 12:2xZ). No file was edited; the branch carries no commits.1. 拆分 PR 的「同一 PR 内整包净减 ≥10%」是算术门槛,还是防增长护栏?
一句话问题:#14296 第 1 项裁定 A 的条件 (a) 写的是「同一 PR 内整包 token 净减 ≥10%」。实测 objectstack-ui 的拆分 PR 做不到:包总量 33,957 tok(SKILL.md 22,856 + 生成物 9,710 + evals 1,391),10% = 3,396;而拆分按条件 (c) 是逐字搬运(rules/ 里每句都必须来自 SKILL.md),本身不减字;第 1 轮 PR #14338 已把审计里所有 HIGH/MED 的删除项落地(−2,585);拆分时还能再删的重述只剩 345 tok,减去指针与标题开销 289 tok,本轮授权面上的最大净减 = −56 tok(−0.16%)。唯一能跨过 10% 的行是生成物
contracts/react-blocks.contract.json(−5,352 = −15.8%),但那是第 3 项裁给 spec 车道的缝卡 #14559,明令 ui 飞行不碰生成物。同一算术也将落在 pm-dispatch 的拆分(#14300 第二轮):第 1 轮 #14444 已合并,拆分同样逐字搬运。选项 × 真实代价
gen:react-blocksstops emittingskills/objectstack-ui/contracts/react-blocks.contract.json— the markdown rendering becomes the single AI-facing output (#14296 item 3 = A) #14559 −5,352 = −7,993,对程序前 36,542 为 −21.9%。每项改动留在自己的车道、由自己的 owner 评审。代价:拆分 PR 本身不再有「≥10%」这条可机械核验的数字,靠 PR 正文的程序级账目代替。gen:react-blocksstops emittingskills/objectstack-ui/contracts/react-blocks.contract.json— the markdown rendering becomes the single AI-facing output (#14296 item 3 = A) #14559 的生成物删除合进一个 PR,字面满足 (a):−5,408 = −15.9%。代价:一个 PR 跨两个域(生成器packages/spec/scripts/build-react-blocks-contract.ts、scripts/regen-artifacts.mjs、scripts/pm/check-governed-merges.mjs、棘轮自检的排除集都随之动),两个 owner 审一个 PR;Seam (spec lane, ruled):gen:react-blocksstops emittingskills/objectstack-ui/contracts/react-blocks.contract.json— the markdown rendering becomes the single AI-facing output (#14296 item 3 = A) #14559 已立卡归domain:spec。gen:react-blocksstops emittingskills/objectstack-ui/contracts/react-blocks.contract.json— the markdown rendering becomes the single AI-facing output (#14296 item 3 = A) #14559 落地后再派拆分。代价:拆分 PR 自身的测量仍是 −56,(a) 在 PR 级仍不成立 —— C 塌缩为 A 加延迟。业务含义:A = 客户的 agent 建列表页只装列表页那一节(入口 ≈3.8k,按需再装 2.1k–6.3k 的一个 rules 文件);D = 每建一个视图都把仪表盘、报表、页面、动作全装一遍。
四维:① 长远合理性 —— rules/ 是目录已定义的结构(README「Skill anatomy」),data(7 个 rules)、platform(4)、query(3)已在其上;ui/pm-dispatch 迁上去是收敛不是增生 —— 与第 1 项裁 A 时的轴 ① 同一论证。「没有 10% 可删」恰恰是第 1 轮做对了的证据;② 实际拉动 —— 不受算术影响:
defineAction34 /definePage29 /defineView10 次真实用法散在四个互不相关的节,今天每次都付 22,856;③ 防 AI 犯错 —— 入口短、相邻无关例子少(第 1 项轴 ③);要申报的一项反向代价:check:skill-identifier-livenessLeg 2 有三条绑定按路径钉在skills/objectstack-ui/SKILL.md(## Report Types、### Action Types、### Navigation Item Types),这三个标题必须留在入口而正文外移,形成三处「一题分两文件」;④ 不扩散 —— A 只加指针不加字;B 让一个 PR 横跨两个域;D 不花钱也不存钱。推荐 A;回退 C(实为 A 加延迟)。置信缺口:「同一 PR」当初由本席位起草、维护者照单采纳 —— 起草意图是轴 ④ 的防增长护栏(「把拆分钉死为瘦身手段」),但字面是逐 PR 算术;哪一种读法有效,只能由维护者定,本席位不重裁。
裁后执行:A ⇒ #14298 第二轮按原简报续飞(同一 dev、同一分支),PR 正文列程序级账目与 −56 的实测,棘轮新增行钉在落地计数、入口行重锁;#14300 第二轮同读法。B ⇒ #14559 并入 ui 拆分,缝卡关闭为 duplicate,PR 加请 spec 车道评审。D ⇒ #14298 / #14300 第二轮关闭为 not planned,棘轮上限各自重锁。
① 长远:收敛到已定义的 rules/ 结构,「无 10% 可删」是第 1 轮成功的证据 · ② 拉动:四类用法散四节、每次付全额,不受算术影响 · ③ 防错:入口短;三条 Leg 2 路径绑定是要申报的反向代价 · ④ 不扩散:A 只加指针 · 推荐 A · 置信缺口:「同一 PR」的字面 vs 起草意图,维护者定
Refs: #14292 (program anchor) · #14296 (batch 1, ruled
1A 2A 3A 4A) · #14298 (the measurement) · #14300 (same question, second split) · #14559 (seam card, item 3).