Repository navigation
Skills catalog optimization program (maintainer-directed): audit every skill on both roots, then optimize — program anchor #14292
Description
Activity
Phase A complete — roster (skills lane seat, session
session_01LraLgQVGq8egUwfYZpbYt1, 2026-09-02 ~01:55Z)All 17 read-only audits are in: 12 published packages, the four small internal skills (+ objectui
verify), and the four PM-corpus classification tables (posted on #13597 as Phase 1 Tables 3–6). Every published package is DIET or RESTRUCTURE; measured reducible mass ≈ 25–35% per package without losing a decision, plus incidental falsehoods the truth sweep did not reach (nine packages). Batch-1 decisions are on #14296 (4 items, recommendations1A 2A 3A 4A).package verdict net Δ (audit) member card state objectstack-data RESTRUCTURE ≈ −10,150 (−22%) #14297 dispatched (flight 1) objectstack-ui RESTRUCTURE ≈ −9,170 (−25%) + −5,352 generated (ruling) #14298 dispatched (flight 1) objectstack-platform RESTRUCTURE ≈ −5,970 (−25%) #14299 dispatched (flight 1) objectstack-pm-dispatch (published) RESTRUCTURE ≈ −4,255 (−29%) #14300 queued objectstack-query RESTRUCTURE ≈ −4,088 (−33%) #14302 queued objectstack-api RESTRUCTURE ≈ −2,125 (−31%) #14304 queued objectstack-ai RESTRUCTURE ≈ −2,303 (−29%) #14305 queued objectstack-upgrade DIET ≈ −1,808 (−22%) #14307 queued objectstack-automation RESTRUCTURE ≈ −1,125 net / −2,105 gross #14308 queued objectstack-i18n DIET ≈ −1,685 (−27%) #14315 queued objectstack-formula DIET ≈ −1,265 (−21%) #14316 queued objectui skills/objectuiRESTRUCTURE ≈ −21,600 (−34%) objectstack-ai/objectui#7251 queued internal small skills (+ #14058 folded) TRIM/DIET/RESTRUCTURE ~85 lines #14317 queued PM corpus (via #13597) dense core, accreted shell AGENTS.md 1162→ 660 · SKILL.md 984→840 · os-dev.md 469→350 · references 1102→891Phase 2 batches follow Phase 1 complete Standing cross-package rulings applied to every flight (recorded here once): the package whose generated
references/_index.mdlists the governing Zod schema as a core schema is the ANCHOR for that topic; every other package deletes its copy and leaves a one-line pointer; a missing or weaker anchor row is a "follow-up for " line in the PR body, never an edit in another package. Concretely: date macros / filter tokens → objectstack-query; objects / fields / relationships / hooks / seeds / datasources / permissions → objectstack-data; plugin wiring,os serveauto-registration, service registry → objectstack-platform; CEL → objectstack-formula; screen flows → objectstack-automation; dataset analytics /compareTo→ objectstack-ui. Eval stubs, new files, generated-file changes and cross-package growth are deferred to #14296.Quota note: the maintainer reported the 5-hour window at 79% (reset ≈ 05:53Z); flights 2+ dispatch after flight 1 lands or after the reset, whichever comes first.
Generated by Claude Code
Closed
completed: the skills catalog optimization program is delivered, and every member card is closedTriage seat (objectstack-wide, seat post #6015) ·
session_01AavokzJ5DndAwitDXvKy4U· 2026-09-28T14:34Z.Provenance: the maintainer's ruling. In the triage seat's chat (2026-09-28), the long-term-program batch 1 was presented in the director format with this card as item ④. Option A was 「按完成关闭」, recommended A with B as the fallback. The maintainer replied, verbatim: 「13597 A 14292 A 13658 A」.
Delivered, read today. Every card named on the Phase A roster (
5503139875, the comment above) is closed:- the decision batch [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;
- the member cards Skills optimization flight — skills/objectstack-data: RESTRUCTURE (≈ −10,150 tok, −22%) — hooks taught twice across two files, generic-JS patterns, 3.3% example gate coverage,
sharingModelused by 31/31 objects and explained nowhere #14297, Skills optimization flight — skills/objectstack-ui: RESTRUCTURE (≈ −9,170 tok, −25%) — a 25.4k-token single entry file, dashboards taught in two sections 640 lines apart, a generated contract shipped twice, and the canonical dashboard example teaching two retired widget keys #14298, Skills optimization flight — skills/objectstack-platform: RESTRUCTURE (≈ −5,970 tok, −25%) — an orphan rules file with zero inbound links, the canonical hooks reference two hops from the entry file, "When to Use" printed twice, andrequires:/engines.protocol/onEnable(every real config) taught nowhere #14299, Skills optimization flight — skills/objectstack-pm-dispatch: RESTRUCTURE (≈ −4,255 tok, −29%) — monorepo-only lore inert in a customer project, the report JSON printed twice, and four drifts from the internal dev contract (report double-channel,premise_still_valid, the stash ban, theSession:line) #14300, Skills optimization flight — skills/objectstack-query: RESTRUCTURE (≈ −4,088 tok, −33%) — seven rules stated twice across SKILL.md and rules/, 27% of tokens on an aggregation surface with zero app-author call sites,context(74 real usages) taught nowhere, and the canonicalexpandexample expands nothing #14302, Skills optimization flight — skills/objectstack-api: RESTRUCTURE (≈ −2,125 tok, −31%) — the leading canonical example teachesRestApiEndpointSchema(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, 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, Skills optimization flight — skills/objectstack-upgrade: DIET (≈ −1,808 tok, −22%) — a stale pinned17.0.0-rc.5registry snapshot (drift 29–165%), one rule stated five times, and a--stored --applycommand that exits 1 for every agent reader because it omits--yes#14307, Skills optimization flight — skills/objectstack-automation: RESTRUCTURE (≈ −1,125 tok net / −2,105 gross) — the package teaches a control-flow trio with zero authorings (parallel_gateway/join_gateway/boundary_event) and never mentions the trio the corpus uses (parallel/try_catch/loop);notify(#1 node type, 20 uses) has one table cell; the description still promises ETL pipelines removed in spec 17 #14308, Skills optimization flight — skills/objectstack-i18n: DIET (≈ −1,685 tok, −27%) — "Core Concepts" restated in full four sections later, a per-namespace layout with zero real usages,os i18n checktaught in six places, and the #1 real pattern (a package shipping its own generated bundle) dead-ends #14315, Skills optimization flight — skills/objectstack-formula: DIET (≈ −1,265 tok, −21%) — 21% of the file onpreviousvsrecord, cron/template surfaces the frontmatter routes away (0 corpus usages), and the one decision the package exists for (formula fields are virtual:where/orderByon one = 400) taught nowhere #14316, Skills optimization flight — the four small internal skills (+ objectuiverify): seven dead references, three rules copied from os-dev.md / SWEEP.md / RUNNER.md, thelive-elsewheregap (#14058), and a versioned Chromium path that breaks on the next image bump #14317 and [finding]spec-property-retirementskill predates thelive-elsewhereverdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058.
The one open reference was #13597, the PM-corpus audit, closed in the same ruling.
Next round. A later optimization round is a new card with its own roster. ⛔ It does not reopen this one.
Program anchor, filed by the skills lane seat (session
session_01LraLgQVGq8egUwfYZpbYt1, GitHubos-litant) on the maintainer's in-session order.Mandate (verbatim, untranslated)
2026-09-02, chat: 「审核所有的 skills，进行全面的优化。」 Standing context: 2026-08-21 「对外发布的 skills 是整个平台的最大价值，尤其要整体考虑和评估。」 · 2026-08-21 「需要整体考虑 skills 的长度，不能为了一个小功能扩写很多。」 · 2026-08-31 「最近我也感觉项目经理越来越笨，是不是胡乱写的文字太多了」.
Relationship to the two prior programs — builds on, does not repeat
skills/**both repos.claude/skills/pm-dispatch/**, os-dev.md, AGENTS/CLAUDE⛔ This program does not re-verify claims one by one; it reads the corpus from the seat of the customer agent that loads a whole package, and from the seat of the repo agent that loads the internal skills.
Scope (measured at objectstack
origin/maina59f78d · objectuiorigin/main67dadd6)skills/skills/.claude/skills/verify.claude/agents/os-dev.md·AGENTS.md·CLAUDE.md(via #13597)Published bundle at pin: 171,406 ratcheted units + 16,448 generator-owned = 187,854 (
ceil(utf8/4),scripts/check-skills-token-ratchet.mjs).Method
Phase A — read-only audit (in flight). One auditor per published package (opus), one for the small internal skills (opus), four contract-tier auditors for #13597 Tables 3–6. Fixed rubric, nine dimensions: A trigger precision · B structure / progressive disclosure · C cross-package redundancy · D density · E constructs & examples (top-5 real authoring tasks derived from
examples/**andpackages/apps/**, not from the doc) · F gaps vs real usage (both directions: uncovered frequent patterns; documented surfaces with zero real usage) · G consistency & retired-spelling residue · H evals · I size proportionality. Every finding carries a span, measured evidence, a proposal (DELETE / MERGE-INTO / MOVE-TO / REWRITE-AS-CONSTRUCT / ADD-paid-by-deletion / RETIRE-SURFACE), a signed token delta, the three value-gate answers, confidence, and a human-floor flag. Verdict scale per package: HEALTHY / TRIM (<10% reducible) / DIET (10–30%) / RESTRUCTURE (>30% or structural).Phase B — cards. Auditor reports are cross-read (twin findings across packages), passed through the standing value gate (① changes a specific decision ② deletion over addition ③ construct over prose), deduplicated against open lane cards, and filed one member card per package linking back here. Findings on the human floor — removal of published capability text, ceiling raises, cross-file restructures, retired/deprecated semantics — go to the maintainer in batches of ≤5 with recommendations, ⛔ never executed on seat judgment.
Phase C — flights. One os-dev flight per package, ≤3 in parallel, no file overlap;
skills/objectstack-pm-dispatch/**and.claude/skills/pm-dispatch/**never in one PR. Token ratchet discipline: byte-neutral-or-shrinking per file, ⛔ no re-wrap as payment, ⛔ no ceiling raise by a dev. Governed posture: any.mddiff stays draft, review-requested to the authorized approvers; pure code faces reviewed in-seat at contract tier. Each PR body carries the per-item 落点 | before | after list; zero out-of-list changes.Roster
Member cards are filed after Phase A per package and listed here as they are filed. Phase-1 Tables 3–6 of the PM corpus post on #13597, not here.
Refs: #13658 (truth sweep, method precedent) · #13597 (PM corpus audit, parallel corpus) · seat post #7623 (running state).