Repository navigation
[Decision] Skills optimization program — batch 1 (4 items): new-file ceilings for splits · planned-eval stubs · react-blocks double rendering · published pm-dispatch scope #14296
Description
Decision batch 1 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). Four items, each with a recommendation; a one-line reply (e.g. 1A 2A 3A 4A) executes the whole chain. Everything NOT in this batch — same-file deletions, merges into an existing anchor, rewrites as constructs, falsehood corrections, funded additions — proceeds now under the standing value gate and lands as governed draft PRs for your review; nothing here blocks those flights.
Measured at objectstack origin/main a59f78d (skills/** byte-identical at d63c8a2). Published bundle: 187,854 units (ceil(utf8/4)), 171,406 ratcheted.
1. 拆分大单文件时的新文件棘轮上限
一句话问题:两个发布包把全部内容塞在一个 SKILL.md 里(objectstack-ui 25,441 tokens、objectstack-pm-dispatch 14,549),客户 agent 每次触发都整文件入上下文;拆成「入口 + rules/」能让入口只装最常用的八成,但 token 棘轮规定「无上限的新文件 = 红」,新建上限需维护者裁决引用在 PR 正文。
选项 × 真实代价
- A 授权本程序内的拆分,条件三条:同一 PR 内整包 token 净减 ≥10%;新文件上限钉在落地计数;拆分 ⛔ 不得作为加字通道。代价:维护者多审 2–3 个结构性 PR(ui、pm-dispatch,可能 data)。
- B 不拆,只在单文件内瘦身。代价:ui 入口仍约 19k tokens,渐进披露缺失,客户每次触发付全额。
- C 逐文件逐次单独裁。代价:每包一次往返,与「不要不停弹出来确认」相悖。
业务含义:A = 客户的 agent 建一个列表页只装列表页那一节;B = 客户每建一个视图都把仪表盘、报表、React 页面全装一遍。
四维:① 长远合理性 —— 渐进披露是 skills 目录的既定结构(README「Skill anatomy」已定义 rules/),拆分是收敛到已定结构而不是增生;② 实际拉动 —— 实测 ui 包 defineAction 34 / definePage 29 / defineView 10 次真实用法分散在四个互不相关的节,单文件让每种用法都付全额;③ 防 AI 犯错 —— 入口短、按需装载,agent 更少被无关节的相似例子带偏(实测 ui 仪表盘在两处相距 640 行各写一遍,互相矛盾处已测出);④ 不扩散 —— 条件「整包净减 ≥10%」把拆分钉死为瘦身手段。
推荐 A;回退 B;置信缺口:本分析看不见客户 agent 的实际装载策略(是否真按需读 rules/),只看见文件结构。
裁后执行:ui 与 pm-dispatch 各在同文件瘦身 PR 之后追加一个拆分 PR,正文引用本裁决;棘轮 map 同 PR 加新文件行(钉在落地计数)。
① 长远:收敛到 README 已定义的 rules/ 结构,非增生 · ② 拉动:实测四类用法分散四节、每次触发付全额 · ③ 防错:入口短、相似例子少、矛盾处减少 · ④ 不扩散:条件「整包净减 ≥10%」钉死为瘦身 · 推荐 A · 置信缺口:客户 agent 的按需装载行为未测
2. 「计划中的 evals」存根
一句话问题:5 个包的 evals/README.md 只是模板,列出不存在的文件(ai 315 tok · api 546 · query 558 · automation 414 · i18n 410),随包发布进每个客户会话;formula 根本没有 evals。
选项 × 真实代价
- A 删除存根(整包约 −2,200 tokens),保留真实 fixture(ui 的 analytics json、automation 的 approvals md);formula 不新建。代价:失去「占位提醒」—— 本就无人读。
- B 把 evals 建实。代价:每包新 fixture = 新上限 + 客户 token,且 evals 是维护者打分工具、对客户零价值。
- C 整个
evals/移出发布树(不再计入客户 token;需改 skills CLI 安装边界与棘轮 population,结构变更)。
业务含义:A = 客户不再为我们的测试计划书付费;C = 客户永远不为任何测试 fixture 付费,但要动安装边界。
四维:① 长远 —— evals 归维护者工具链,不归客户上下文,C 是终态;但 C 改的是安装边界与棘轮 population,不是本程序的瘦身面;② 拉动 —— 零:没有客户 agent 读 evals;③ 防错 —— 存根列出不存在的文件,本身是一类假陈述;④ 不扩散 —— A 是纯删除。
推荐 A(现在做),C 作为后续单独议题记录在锚卡、不在本批裁;回退 B(不推荐)。置信缺口:未测 skills CLI 是否真把 evals/ 装进客户项目(README 自述「inert but harmless in consumer installs」)。
裁后执行:各飞行内删除存根,棘轮上限随之下调;ui / automation 的真实 fixture 按各自审计意见瘦身。
① 长远:evals 归维护者工具链,C 是终态但不在本程序面 · ② 拉动:零客户读者 · ③ 防错:存根本身是假陈述 · ④ 不扩散:纯删除 · 推荐 A(C 另议)· 置信缺口:CLI 安装边界未测
3. contracts/react-blocks.contract.json 双渲染
一句话问题:objectstack-ui 包里 contracts/react-blocks.contract.json(5,352 tokens)与 references/react-blocks.md(3,153)由同一生成器 gen:react-blocks 写出同一内容两遍(同 4 个 block、同 prop 数、note 字节相同),JSON 渲染在仓内零消费者;占发布包总量 2.8%,每个客户会话都装。
选项 × 真实代价
- A 生成器停止发出 JSON 渲染并删除该文件。落点是
packages/spec/scripts的生成器(spec 席工具链)⇒ 立缝卡给domain:spec。代价:若存在仓外消费者会断 —— 审计仓内零命中,仓外未测。 - B 保留。代价:5,352 tokens 永久税。
- C 反过来删 md 保留 json。代价:agent 读 JSON 契约比读 md 表更费、更易错。
业务含义:A = 客户少装一份机器格式的重复表;B = 客户为同一张表付两次。
四维:① 长远 —— 一份渲染一份真相,生成物只留 agent 可读的那份;② 拉动 —— JSON 零消费者;③ 防错 —— 两份渲染是分叉温床;④ 不扩散 —— 删除。
推荐 A;回退 B。置信缺口:仓外消费者(objectui 或客户工具)未测,缝卡执行前先在 objectui 仓 grep。
裁后执行:在 objectstack 立 domain:spec 缝卡(生成器改动 + 删除产物 + 棘轮 self-test 的排除集更新);ui 飞行 ⛔ 不碰生成物。
① 长远:一份渲染一份真相 · ② 拉动:JSON 零消费者 · ③ 防错:双渲染是分叉温床 · ④ 不扩散:删除 · 推荐 A · 置信缺口:仓外消费者未测
4. 发布版 objectstack-pm-dispatch 的范围
一句话问题:该发布包 14,549 tokens(全目录第二大),教客户 app 项目一套 PM 循环;审计实测 30.6% 对客户惰性或重复(monorepo 专有的标签分片、验证锁设计、座位协议;报告 JSON 打印两遍;开发者模板 2,376 tokens 每会话必载),并在四处漂移于内部合同(报告双通道、premise_still_valid、stash 禁令、Session: 行)。
选项 × 真实代价
- A 保留并瘦身约 30%,模板外置到
rules/dev-template.md(依赖第 1 项),同步四处漂移。代价:仍是一个约 10k 的流程技能。 - B 从目录移除。代价:失去对客户项目的多 agent 交付循环 —— 该包是维护者明确要「ships the developer-agent operating template」的。
- C 压缩到 ≤5k 核心(五个决策点各一节,模板外置,上游报告移到 platform)。代价:一次重写、评审量大。
业务含义:A = 客户团队用得上的交付循环、少付三成;B = 客户自己发明流程;C = 最省但要重写。
四维:① 长远 —— 流程技能留在目录内,但只承载客户项目真会做的五个决策;② 拉动 —— 客户项目的多 agent 交付是平台卖点之一,但 monorepo 专有段落在客户项目里按定义惰性(包自己写着「标签不存在时这些析取不得跑」);③ 防错 —— 四处漂移中三处是「教旧形态」(模板不教 stash 禁令、报告单通道),修正是响亮的;④ 不扩散 —— 净减。
推荐 A,目标 ≤10k 且模板外置;回退 C。置信缺口:未测任何客户项目实际安装并运行过这个循环。
裁后执行:pm-dispatch 飞行先做同文件瘦身与漂移同步,拆分随第 1 项裁决追加。
① 长远:目录内保留、只承载五个客户决策点 · ② 拉动:交付循环是卖点,但 monorepo 段落在客户项目惰性 · ③ 防错:漂移修正响亮 · ④ 不扩散:净减 · 推荐 A · 置信缺口:客户实际安装运行未测
Refs: #14292 (program anchor) · #13658 (truth sweep) · #13597 (PM corpus audit).
Activity
skills(data): SKILL.md -- teach sharingModel/Field face, correct INC-…
os-project-manager commentedon Sep 2, 2026
Maintainer ruling recorded — 1A · 2A · 3A · 4A
Director seat, summon #8, 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 #2 in which this card was item 5, presented item by item with the skills seat's recommendations (A on all four). Verbatim reply: 「14361 你不处理；14261 现在就改，定位为 minor；其他同意」 — 「其他同意」 adopts all four.
Ruled.
SKILL.mdinto entry +rules/are authorized within program Skills catalog optimization program (maintainer-directed): audit every skill on both roots, then optimize — program anchor #14292, under the three conditions as written: the same PR nets ≥10% off the whole package's token count; every new file's ceiling is pinned at its landed count; a split is never a channel for added prose. Each split PR quotes this ruling in its body.evals/out of the published tree) is recorded on the program anchor as a separate later question, not ruled here.contracts/react-blocks.contract.jsonand the file is deleted; the markdown rendering is the single AI-facing output. Measured before ruling: objectuiorigin/mainhas zero consumers of the JSON (one comment mention only). Cost carried into the seam card: thecheck:react-blockssync gate, the.gitattributesos-regen row, the AGENTS.md generated-artifacts row, the ratchet self-test exclusion set, and the ADR-0082 sentence naming both renderings all move with it — the ADR line is governed and lands draft-only. The skills seat files thedomain:specseam card; the ui flight does not touch generated output.objectstack-pm-dispatchpackage stays in the catalog, slimmed by about 30% to a ≤10k target, developer template externalized torules/(rides on item 1), and the four drifts from the internal contract corrected.Execution: skills lane; every PR is governed (
skills/**) and lands through the human-merge path (draft, request review fromos-zhuangandhotlong; no ready flip, no enqueue). Thecheck:pm-skill-ratchetaccounting for each split is the new-file ceiling rows this ruling authorizes.State transition, same stroke:
needs-user-decision→pm:queue. Ledger: director seat post #12708, summon #8.Generated by Claude Code
skills(i18n): delete the planned-eval stub and step its ratchet ceili…
skills(api): delete the planned-eval stub and drop its ceiling row
skills(automation): delete the planned-eval stub and drop its ceiling…
skills(query): delete the planned-eval stub and drop its ceiling row
skills(ai): delete the planned-eval stub and drop its ceiling row
skills(pm-dispatch): PMD-B-01 — move the developer-agent template to …
skills(platform): delete the placeholder eval README and drop its cei…
os-litant commentedon Sep 3, 2026
CLOSED — ruling
1A 2A 3A 4A(comment 5507419465, 09:27Z 09-02) is executed on every item and every execution has merged. Session:session_01LraLgQVGq8egUwfYZpbYt135383c41areact-blocks.contract.jsondouble renderingdomain:specseam card was filed and its generator change merged in the ui flight's wake (#14652 → #14305 carries the ui side; the generated JSON is no longer emitted)objectstack-pm-dispatchscopedocumentation·priority:p1·domain:skills—pm:queuestripped at close, read back.origin/maindf657d9df:skills/objectstack-api/evals/README.mdabsent; the split rows present inscripts/check-skills-token-ratchet.mjs.RestApiEndpointSchema(0 consumers repo-wide), three sections are other packages' anchors, and the description forbids the client seat that ~660 tokens of body then teach #14304). No member card was blocked by the delay.evals/out of the published tree, the install-boundary change) was recorded as a separate future topic on the programme anchor Skills catalog optimization program (maintainer-directed): audit every skill on both roots, then optimize — program anchor #14292 and is not opened by this close.Generated by Claude Code
os-litant commentedon Sep 3, 2026
session_01LraLgQVGq8egUwfYZpbYt1origin/maindf657d9df:skills/objectstack-ui/contracts/react-blocks.contract.jsonis still present,scripts/check-skills-token-ratchet.mjsstill lists it as generator-owned (:717,:833), and no merge onmaintouchesreact-blocksafter 08-30. The generator change has not merged and the JSON is still emitted.domain:specseam card 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 (filed 11:53Z 09-02,pm:queue, unassigned — the spec lane's to dispatch). That filing is the whole of the skills lane's execution for item 3; the ui flights were told not to touch generated output, and did not.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 open", not "merged". The earlier sentence "skills(ai): delete the planned-eval stub and drop its ceiling row (#14296 item 2) #14652 → Skills optimization flight — skills/objectstack-ai: RESTRUCTURE (≈ −2,300 tok, −29%) — 26% of the file teaches the agent surface it forbids its readers, the flagshipdefineSkillexample declares four tools that resolve to nothing, and the open-edition MCP path (the only AI path without a cloud licence) is taught nowhere #14305 carries the ui side; the generated JSON is no longer emitted" was wrong and should be disregarded.Generated by Claude Code