Skills optimization flight — skills/objectstack-formula: DIET (≈ −1,265 tok, −21%) — 21% of the file on `previous` vs `record`, cron/template surfaces the frontmatter routes away (0 corpus usages), and the one decision the package exists for (formula fields are virtual: `where`/`orderBy` on one = 400) taught nowhere

[os-litant]

Member card 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). Read-only audit at objectstack origin/main a59f78d. The auditor's harness refused to write a findings file; the full findings are recorded in this card and in the audit-record comment the dev posts at claim time (copy of the seat's audit message, provided in the dispatch brief).

Audit summary

SKILL.md 551 lines / 6,002 tok (headroom 0) + generated _index.md 302; no rules/, no evals/ (the only domain package without one). Heaviest sections: §5 previous vs record :286-375 = 1,241 tok (20.7%); stdlib :153-216 = 913; template surfaces + quickref = 511; cron surfaces + quickref = 228. Real usage: P\`` 37 · cel`85 ·F`0 ·cron`0 ·tmpl`0; top-5 tasks: seed dynamic dates (76, well covered), field conditional rules (63, roots incomplete), validation predicates (10, well covered), flow/trigger conditions (5, scope taught WRONG), date-macro filter tokens (21, zero mention).type: 'formula' fields: 0 live usages (the only hits are a removal test). Stdlib: 35 documented = 35 implemented (CEL_STDLIB_FUNCTIONS`), pinned one direction only. Verdict DIET (net −1,265 = −21.1%; structure sound).

Top findings

id	span	proposal	Δ
FRM-F-06 / A-01	:432-445,:446-481,:482-497,:498-513	cron + template surfaces: frontmatter :9 says "Do not use for cron"; 0 corpus usages of cron\``/cronExpression/tmpl``/any formatter → ONE schema-pointing row each (live surfaces: never delete the last mention), delete the quickrefs	−435
FRM-D-05	:328-330,:342-348,:350-358,:370-372	REWRITE-AS-CONSTRUCT — :350-358 restates table row :311 at 165 tok; the rest is cost narration	−310
FRM-D-04	:317-326	DELETE pre-17 logger.warn history; rule complete at :313-317	−144
FRM-B-06	:38-45	MERGE-INTO :535-547; ADR-0032 stated 3× and :536 cross-refs itself	−119
FRM-F-01	:414 (gap)	ADD — a formula field has no column; where/orderBy/searchableFields on one = 400 INVALID_FIELD (17.engine-find-formula-filter-refused.ts:20-45; examples/app-todo/test/derived-flag-removal.test.ts)	+90
FRM-B-05	:514-526	MERGE-INTO :269-285	−81
FRM-D-01 / B-07	:27-31	DELETE "Strategic context"	−74
FRM-D-07	:84-88	DELETE "you do not learn three languages"	−72
FRM-D-02	:90-95	DELETE M9.7 roadmap	−69
FRM-B-03	:364-368	MERGE-INTO :117-144 — has()-on-declared-field rule stated twice	−68
FRM-F-03	:415-416 (gap)	ADD — form-view visibleWhen is client-side only and faults OPEN (view.zod.ts:2353; form-predicate-root-policy.ts:14-19); the package teaches the opposite globally at :41-42, :313	+60
FRM-D-03 / G-03	:527-534	DELETE "Open questions"; CelExprSchema exists nowhere in the repo	−55
FRM-G-05	:421,:472,:508-510	DELETE visibleOn/visibility row + dedupe titleFormat notes	−44
FRM-B-02	:548-551	pointer at node_modules/@objectstack/formula/ which ships dist only (:154-155 itself calls it unreadable)	−48
FRM-F-02 + F-04	:100-113, :382	ADD roots parent (field.zod.ts:800-801; 4 usages in invoice.object.ts) and data (form-predicate-root-policy.ts:181-189); flow conditions bind bare identifiers (service-automation/src/engine.ts:7150-7156; 5 showcase flows) — the translation table implicitly forbids it	+60

Also: FRM-A-02 (+18: say this is the companion package four hosts load alongside) · FRM-B-01 (+30: link references/_index.md — the only domain package that does not) · FRM-B-04 ("no js dialect" 3×, −36) · FRM-D-09 (:64-71 transcribes expression.zod.ts:75-90, −41) · FRM-E-01 (mark one "Mandatory pattern" block os:check, +45) · FRM-E-03 (:241 teaches * 100.0, a workaround stdlib.ts:246-275 closed) · FRM-F-07 (:429-430 forbids CEL in view filters without naming the date-macro alternative → pointer to objectstack-query) · FRM-G-01 (objectstack build 5× vs os build 5× — siblings use os) · FRM-G-02 (frontmatter 17.x while the body says 7.6, M9.5, M9.7, M9.8) · FRM-G-04 (:105 "single quotes (always)" vs the corpus writing "…" inside P\`` 10+ times). Incidental: skills/README.md:100"LegacyOLD/NEWevaluate tonull since M9.5" is stale — current truth is a build failure (validate.ts:342-352`) — "follow-up for skills/README.md".

Flight scope

IMPLEMENT (same-file, shrink-only): every row above; the three funded additions (F-01 paid by D-01 + D-03; F-03 paid by A-01's deleted quickrefs; F-02 + F-04 paid by D-05); FRM-A-02, B-01, E-01, E-03, F-07, G-01, G-02, G-04.

ANCHOR RULINGS: this package is the CEL anchor; host packages (automation :41-43,:943-946; data :1111-1118) restate the stdlib beside pointers here — record under "follow-up for objectstack-automation / objectstack-data" (⛔ do not edit them). Date-macro filter tokens → objectstack-query (pointer only).

DEFER: FRM-H-01 (a new evals/README.md needs a ceiling row → #14296 item 2 / a later ruling); FRM-C-05 (generator SKILL_MAP maps date-macros.zod.ts here, contradicting :36-37 → out-of-scope card for the spec lane); FRM-D-06 (moving the 10,000-row limit to a host package → follow-up); FRM-I-01 ceiling re-lock (Phase 3 style, not this PR).

Flight constraints (binding)

• ONE draft PR, first line Fixes #<this card>; governed ⇒ stays draft; review requests are the seat's step.
• Token ratchet: SKILL.md may not grow; additions paid by deletions in the same file; ⛔ re-wrap is not payment; ⛔ no ceiling raise; ⛔ no new files; ⛔ do not touch the ratchet script.
• ⛔ Never edit another package's files or skills/README.md prose; follow-ups in the PR body. Generated files untouched.
• Live surface with zero measured usage ⇒ one row pointing at its schema; retired/tombstoned ⇒ delete. A false claim matching a spec .describe() string ⇒ spec-side twin card.
• Gates: node scripts/check-skills-token-ratchet.mjs, pnpm --filter @objectstack/spec check:skill-examples, pnpm check:skill-compatibility, pnpm check:skill-identifier-liveness, pnpm --filter @objectstack/spec check:skill-docs, plus node scripts/pm/dispatch-gates.mjs --commands skills/objectstack-formula/SKILL.md; record the head sha. Note: packages/formula pins the stdlib catalog ↔ skill in cel-stdlib-drift.test.ts and skill-catalog-sync.test.ts — run them.
• PR body: per-item 落点 | before | after keyed by finding id; token delta; needs:contract-review on both carriers (F-01, F-03, F-04 are contract-semantics claims).

Refs: #14292 · #14296 · #13658.

[os-litant]

Claim: PM loop round 3
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14316-skills-formula-optimization
Worktree: objectstack-issue-14316
Domain: domain:skills
File surface: skills/objectstack-formula/SKILL.md ONLY (shrink-only; ceiling 6,002, headroom 0); generated references/_index.md untouched; ⛔ no new evals/ file (FRM-H-01 deferred, #14296 item 2); skills/README.md + content/docs/ai/skills-reference.mdx regenerated only if check:skill-docs demands it after the FRM-A-02 description edit (⛔ no hand edit to README prose — the stale "Legacy OLD/NEW evaluate to null" line at skills/README.md:100 goes in the PR body as a follow-up); scripts/role-word-baseline.json only if check:role-word ratchets DOWN and names --update as the remedy (its own commit); ⛔ no other package (stop on breach; explain in the report)
Container & model: S/M (one 6,002-token file, 551 lines; three funded additions with source-pinned semantics), mode:subagent, model: opus — --tier reads no path-derived mandate (floor sonnet · default opus · ceiling fable); the content limb applies, compensated by the seat's in-seat review at CONTRACT_REVIEW_TIER
Clause-②: yes — FRM-F-01 (a formula field has no column: where / orderBy / searchableFields on one = 400 INVALID_FIELD), FRM-F-03 (form-view visibleWhen is client-side only and faults OPEN), FRM-F-02 / F-04 (the parent / data roots; flow conditions bind bare identifiers) are contract-semantics claims ⇒ needs:contract-review on both carriers at PR creation; in-seat review before any human review request
Serial constraints cleared: this package is the CEL ANCHOR — the delivered automation flight (PR #14467) and the query flight (PR #14443) deleted their CEL restatements on the ruling that formula's stdlib table, the OLD/NEW translation row and its warning, and the CEL-carrier table are the anchors, so those three constructs must survive this diet with their headings addressable; date-macro filter tokens → objectstack-query rules/filters.md (pointer only); no open PR touches skills/objectstack-formula/**; in flight alongside #14315 (i18n) on a disjoint directory; #14302 / #14307 / #14308 delivered and under seat review, #14305 in patch round 1

Decision re-read: no maintainer ruling on this card; binding inputs = the mandate, the value gate, the ratchet discipline, and #14296 (item 2 deferred: no eval file). FRM-C-05 (generator SKILL_MAP maps date-macros.zod.ts here) goes out as an out-of-scope finding card, not a doc edit. Landing form: first line Part of #14316 (the deferred eval keeps the card open).

---

Generated by Claude Code

[os-litant]

Claim: dev agent (os-dev), dispatched by the skills lane seat
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14316-skills-formula-optimization (pushed empty at claim time — write-route probe green)
Worktree: objectstack-issue-14316
Base: 93940d49 (origin/main has moved since the dispatch brief was written, which named 18d816a5; skills/objectstack-formula/** is byte-identical across a59f78d, 18d816a5 and 93940d49 — git diff --stat empty both ways, so PM mechanism assumption 1 holds and every skills/ span in the audit is valid at my base. packages/** line numbers are re-cited at 93940d49.)
File surface: skills/objectstack-formula/SKILL.md ONLY by hand (551 lines / 24,006 bytes = 6,002 tok, ceiling 6,002, headroom 0); skills/README.md / content/docs/ai/skills-reference.mdx only as generator output if check:skill-docs demands it after the FRM-A-02 description edit; scripts/role-word-baseline.json only if check:role-word ratchets DOWN and names --update.
Landing form: draft PR, first line Part of #14316 (no closing keyword anywhere); needs:contract-review on both card and PR at PR creation.

Not touching the assignee. Audit record follows as the next comment.

---

Generated by Claude Code

[os-litant]

audit-record — skills/objectstack-formula (read-only audit at a59f78d)

Transliteration key (the GitHub body sanitizer eats angle-bracket-shaped fragments, so every one is spelled out here):

• LT = an opening angle bracket (less-than sign)
• GT = a closing angle bracket (greater-than sign)
• so LT!-- os:check --GT in FRM-E-01 is the HTML-comment-shaped marker os:check that check:skill-examples looks for above a fence.
• && is an upstream escaping artifact carried over verbatim from the auditor's message; it stands for the CEL logical-and operator.

Provenance: this is the auditor's own message, posted verbatim apart from the two substitutions above and the dropped first paragraph (a note about the auditor's harness refusing to write a findings file). Posted by the dev seat at claim time; the dev's base is 93940d49, and git diff --stat a59f78d 93940d49 -- skills/objectstack-formula is empty, so every skills/ span below holds unchanged at the implementation base.

---

Ref note: the checkout is at a39b02a, one commit behind the pinned a59f78d. git diff a39b02a a59f78d -- skills/ is empty and the 5 changed files touch none of packages/spec, packages/formula, examples, packages/apps — every span below holds at the pinned ref.

---

objectstack-formula — audit

Package summary

SKILL.md 551 lines / 6002 tokens (ceiling 6002, headroom 0) + generated references/_index.md 302 tokens = 6304 shipped. No rules/, no evals/.

Heaviest sections (measured ceil(bytes/4)): §5 previous vs record :286-375 = 1241 (20.7%); stdlib :153-216 = 913; Core contract :60-96 = 458; Template surfaces+quickref :446-481,:498-513 = 511; has() :117-144 = 356; CEL surfaces :407-431 = 285; translation table :376-406 = 276; Cron surfaces+quickref :432-445,:482-497 = 228.

Real-usage corpus (examples/** + packages/apps/**, -g '*.ts'): P` 37 · cel` 85 · F` 0 · cron` 0 · tmpl` 0. Top-5 tasks by count: (1) seed dynamic dates cel`daysAgo/daysFromNow/now` 76 — well covered §4; (2) field conditional rules 63 — covered but roots incomplete; (3) validation-rule predicates 10 — well covered (examples/app-showcase/src/data/objects/project.object.ts:114 matches the skill's canonical RIGHT example verbatim); (4) flow/trigger conditions 5 — scope taught wrong; (5) date-macro filter tokens 21 — zero mention. type: 'formula' fields: 0 live usages (2 hits, both in examples/app-todo/test/derived-flag-removal.test.ts, a removal test).

TOP findings (ranked by Δtokens × confidence)

id	span	proposal	Δ
FRM-F-06	:432-445,:446-481,:482-497,:498-513	RETIRE-SURFACE cron (frontmatter :9 routes it away) + MOVE-TO formatter table; 0 corpus usages of cron`/cronExpression/tmpl`/any formatter	−435
FRM-D-05	:328-330,:342-348,:350-358,:370-372	REWRITE-AS-CONSTRUCT — :350-358 restates table row :311 at 165 tok; :342-348/:370-372 are cost narration	−310
FRM-D-04	:317-326	DELETE pre-17 logger.warn history; rule complete at :313-317	−144
FRM-B-06	:38-45	MERGE-INTO :535-547; ADR-0032 stated 3× and :536 cross-refs itself	−119
FRM-F-01	:414 (gap)	ADD — a formula field has no column; where/orderBy/searchableFields on one = 400 INVALID_FIELD	+90
FRM-B-05	:514-526	MERGE-INTO :269-285	−81
FRM-D-01 / B-07	:27-31	DELETE "Strategic context" (engine-choice history, first content after H1)	−74
FRM-D-07	:84-88	DELETE "you do not learn three languages"	−72
FRM-C-01	-automation:41-43,:943-946	DELETE two truncated stdlib restatements sitting beside pointers to this package	−72 (host)
FRM-D-02	:90-95	DELETE M9.7 roadmap; sentence 1 dups :86-88	−69
FRM-B-03	:364-368	MERGE-INTO :117-144 — has()-on-declared-field rule stated twice	−68
FRM-F-03	:415-416 (gap)	ADD — form-view visibleWhen is client-side only and faults open (true)	+60
FRM-D-03 / G-03	:527-534	DELETE "Open questions"; CelExprSchema exists nowhere in the repo	−55
FRM-C-02	-data:1111-1118	DELETE stdlib subset + determinism restatement	−92 (host)
FRM-G-05	:421,:472,:508-510	DELETE visibleOn/visibility row + dedupe titleFormat notes	−44

Remaining measured findings: FRM-A-01 frontmatter :9 says "Do not use for … cron schedules" while the body teaches cron (subsumed in F-06). FRM-A-02 :3-10 — four host packages end with "load objectstack-formula alongside"; this description never says it is a companion (ADD +18). FRM-B-01 references/_index.md is an orphan — 8 of 9 domain packages link theirs, formula is the only 0 (ADD link +30). FRM-B-02 :548-551 points at node_modules/@objectstack/formula/, which ships dist only (packages/formula/package.json) and which :154-155 itself calls unreadable (−48). FRM-B-04 "no js dialect" 3× (:81-83,:476-478,:51) (−36). FRM-C-05 build-skill-references.ts:165 maps date-macros.zod.ts to this package, which contradicts it at :36-37 (−40 generated). FRM-D-08 :173 is 157 tokens of prose carrying 5 rules — table, not prose (≈0). FRM-D-09 :64-71 transcribes expression.zod.ts:75-90 (−41). FRM-E-01 2 of 13 ts fences carry LT!-- os:check --GT, and both are the cron/template blocks — every "Mandatory pattern" is untype-checked (ADD one marked block +45). FRM-E-02 swept all 11 unmarked fences for rot per check-skill-examples.ts:26-42 — none found; F/P/cel/tmpl/cron all export from packages/spec/src/index.ts:263-267. FRM-E-03 :241 teaches * 100.0, the workaround stdlib.ts:246-275 closed. FRM-F-02 :100-113 omits roots parent (field.zod.ts:800-801; 4 usages in invoice.object.ts:238,279,285) and data (form-predicate-root-policy.ts:181-189) (+32). FRM-F-04 :382 teaches bare_field → record.bare_field universally, but service-automation/src/engine.ts:7150-7156 spreads variables to top level "so bare identifiers like status / previous.status resolve — the natural authoring style for record-change start conditions", and 5 showcase flows author them that way (+28). FRM-F-05 stored-field expression: (2 usages) — already in -ui:124-125, no gap here. FRM-F-07 :429-430 forbids CEL in view filters without naming the date-macro alternative. FRM-G-01 mixes objectstack build (5×) with os build/os validate (5×); siblings are near-uniform on os. FRM-G-02 frontmatter says 17.x while the body says 7.6, M9.5, M9.7, M9.8 (M9 4× here vs 1× in the whole rest of the catalog). FRM-G-04 :105 prescribes single quotes "(always)"; the corpus writes "…" inside P` 10+ times. FRM-I-01 ~1,900 tokens serve the 5 real patterns; ~4,400 do not.

Dimension H — is the missing evals/ a gap?

Yes, and it is the sole unexcused omission. All eight other domain packages ship one (-ai 315, -api 546, -automation 414 + 1329, -data 143, -i18n 410, -platform 514, -query 558, -ui 289 + 1102). The only other two without — -pm-dispatch, -upgrade — are both domain: process, which skills/README.md explicitly exempts ("carries SKILL.md alone"); formula is domain: expression.

A proportionate eval is ~350-500 tokens / 6-7 rows (calibrate to -query's 558/9; -data's 143 placeholder is the floor), each testing a decision only this package changes: (1) null-guard vs has() — must emit record.x != null && …, fail a has()-only guard (publish gate, :135-143); (2) Salesforce → CEL — &&/in [...], not AND/IN (...); (3) transition condition on after* via previous.x != record.x, fail ISCHANGED() and fail previous on beforeInsert; (4) seed date cel`daysFromNow(45)`, fail new Date(); (5) coalesce(…, '') on nullable string concat; (6) daysBetween(today(), record.due), fail record.due - today(); (7) once FRM-F-01 lands — "users must filter by is_overdue" → stored column, fail a formula field. ⚠️ Do not write a row that fails a bare identifier in a flow condition — FRM-F-04 shows the platform binds those.

Human-floor list

FRM-A-01/F-06 (removes published cron+template capability text) · FRM-B-06, D-04, D-05 (restructure §5 / callout order) · FRM-D-01, D-03, G-02, G-03 (retired milestone + nonexistent-symbol semantics) · FRM-D-06 (moves the 10 000-row limit to a host package) · FRM-G-05 (retired/deprecated aliases) · FRM-C-01/C-02/C-03 (edit other packages' published text) · FRM-C-05 (generator SKILL_MAP change) · FRM-I-01 (ceiling 6002 → 4737) · FRM-H-01 especially: scripts/check-skills-token-ratchet.mjs:93-100 — "a discovered AUTHORED file with no ceiling is RED" — so a new evals/README.md needs a maintainer ceiling row, and no same-file deletion can pay for it.

Incidental falsehoods

1. Stdlib — no drift, both directions checked as asked. CEL_STDLIB_FUNCTIONS (packages/formula/src/validate.ts:695-706) = 35 names; SKILL.md:157-208 documents exactly the same 35. The claim at :154-157 that it is "pinned by two tests" is accurate (cel-stdlib-drift.test.ts:83-153, skill-catalog-sync.test.ts:22-33). One asymmetry, not a falsehood: only catalog → skill is pinned; a function documented here but absent from the catalog would ship silently. Fixable in the test file at zero ratchet cost.
2. skills/README.md:100 is stale (adjacent file): "Legacy OLD / NEW evaluate to null since M9.5." Current truth is a build failure — packages/formula/src/validate.ts:342-352, and this package's own :210-215. Same line carries the retired M9.5; :102 carries "v5.0 vocabulary".

Verdict — DIET

Deletions −1,554, additions +289, net −1,265 = −21.1% of 6002 (ceiling 6002 → 4737), plus −118 in -automation, −92 in -data, −40 generated. Structure is sound (single file, correct 80%), so this is not RESTRUCTURE — but the package is simultaneously oversized for what customers author and missing the one decision it exists to inform.

3 highest-value additions, each with its paying deletion

1. FRM-F-01 — formula fields are virtual (+90), paid by FRM-D-01 + D-03 (−129). packages/spec/src/migrations/entries/semantic/17.engine-find-formula-filter-refused.ts:20-45: "formula is the one field type no driver materialises a column for … a where on a formula field cleared every gate … answered 200 with zero rows"; both doors now 400 INVALID_FIELD. examples/app-todo/test/derived-flag-removal.test.ts:20-30 is a whole app reversing on this. Catalog coverage today: the search axis 3× (-data:141-150, -ui:407, -query:528); the filter and sort axes nowhere.
2. FRM-F-03 — form-view predicates fail open (+60), paid by FRM-A-01 (−228). packages/spec/src/ui/view.zod.ts:2353: "CLIENT-SIDE only: nothing server-side evaluates a form-view field visibleWhen, so a role test here hides the control and protects no data … the root is unbound and the predicate faults open." metadata-core/src/form-predicate-root-policy.ts:14-19 adds "visibleWhen's fault fallback is true" with a measured production incident. This package teaches the opposite globally (:41-42, :313), and visibleWhen is in its own frontmatter.
3. FRM-F-02 + F-04 — the two missing roots and the flow-condition scope (+60), paid by FRM-D-05 (−310). parent (field.zod.ts:800-801, 4 real usages) and data (form-predicate-root-policy.ts:181-189) are absent from the cheat-sheet; and service-automation/src/engine.ts:7150-7156 binds bare identifiers in flow conditions, which 5 showcase flows rely on and the translation table implicitly forbids.

---

Generated by Claude Code