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, `sharingModel` used by 31/31 objects and explained 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 (skills/** byte-identical at d63c8a2); 43 findings DATA-A-01 … DATA-I-04. The full findings table is the audit record: the dev posts it verbatim as the first comment on this card at claim time (source: the seat container's scratchpad audit/objectstack-data/findings.md).

Audit summary

12 files, 4,942 lines, 45,780 hand-authored tokens (+1,483 generated) = 25.2% of the published catalog, 29% larger than the next package; total headroom across 10 ratcheted files: 47 tokens, five at 0. Real usage (examples/** + packages/apps/**): 31 ObjectSchema.create() calls in 26 object files, ~250 Field.* calls vs 22 literals (92% factory), 3 defineSeed modules, 2 hook files, 2 datasource files, 0 *.seed.ts. All 9 patterns with ≥3 real usages are covered; 8 documented surfaces have zero corpus usage; two high-frequency keys have no coverage at all. Verdict RESTRUCTURE: hooks (14,779 tok, 32% of the package) are split across rules/hooks.md and references/data-hooks.md where the second calls itself a pointer and then restates five sections; the rules/ routing table is buried at line 312 of 1,216; the package holds 41% of the catalog's TypeScript (151 fences) with 3.3% os:check coverage — seven unmarked self-contained blocks would fail check-skill-examples today.

Top findings (token-delta × confidence)

id	span	proposal	Δtok
DATA-B-01 / I-03	rules/hooks.md (whole) ∥ references/data-hooks.md	MERGE-INTO one file — hooks.md:3 calls itself "a reference pointer", then restates 5 sections	−1,880
DATA-D-04	references/data-hooks.md:765-1046	DELETE patterns 1,2,4,5,7,9 (generic JS); keep 3,6,8,10	−1,000
DATA-C-05	SKILL.md:858-971	REWRITE-AS-CONSTRUCT — 3 near-identical examples + Studio-banner narration for a 4-value enum; 0 corpus uses	−600
DATA-B-04	references/data-hooks.md:1050-1133	DELETE — "Registration, three methods" also at rules/hooks.md:145-176	−566
DATA-D-11	rules/naming.md:15-93, 103-107	DELETE — 6 ❌/✅ pairs + rationale for a rule the regex at :9 already states	−566
DATA-D-06	rules/indexing.md:161-234	REWRITE-AS-CONSTRUCT — one textbook fact (left-to-right) in 3 subsections	−480
DATA-B-03	rules/hooks.md:99-126 ∥ data-hooks.md:41-70	DELETE one copy of the 8-event table	−466
DATA-C-04	25 tombstone sites	DELETE standalone blocks — each key is refused by a parse error carrying its own prescription; objectstack-upgrade owns this	−450
DATA-C-03	references/data-hooks.md:1209-1275	DELETE — vitest + LiteKernel harnesses; platform's frontmatter claims that surface	−432
DATA-D-07	rules/indexing.md:277-298, 312-340	DELETE — unsourced Table Size → Max Indexes table, pg_stat_user_indexes SQL, 10 best-practices	−430
DATA-B-05	references/data-hooks.md:1136-1160	DELETE — DO/DON'T also at rules/hooks.md:180-194	−418
DATA-D-12	SKILL.md:540-555	DELETE — "CRM Schema Blueprint" points at 6 things covered above it	−358
DATA-C-01	SKILL.md:162-189	DELETE — two verbatim CLI/HTTP error transcripts; the 400 text is word-for-word in objectstack-query	−343
DATA-D-09	rules/validation.md:420-438	DELETE — best-practices + unsourced perf claims	−338
DATA-D-01/02/03	references/data-hooks.md:1279-1444	DELETE — invented benchmark table, "Advanced Topics" (a for loop), Troubleshooting, Summary	−990

Net after +825 of paid additions ≈ −10,150 tokens (−22.2%), touching none of the 9 live patterns.

Incidental falsehoods (fix in this flight)

• INC-01 SKILL.md:393 — required: true → false "dev auto-heals (autoMigrate:'safe')". Nullability drift is compared against storage.notNull, NOT required (packages/drivers/driver-sql/src/schema-drift.ts:795-798; field.zod.ts required.describe()), and the emitted op is severity: 'warning', category: 'needs_confirm', op: 'relax_not_null' — never auto-applied. HIGH.
• INC-02 SKILL.md:398-401 — the "stale NOT NULL column" 400 story: schema-drift.ts:800-802 says that configuration produces a raw driver error, not a clean validation 400. MED.

Three funded additions

1. sharingModel (DATA-F-01, +130, paid by DATA-C-05) — used by 31/31 real objects, absent from both property tables; an object publish with no authored sharingModel is refused (security-owd-unset, packages/lint/src/validate-security-posture.ts:62).
2. Field.* factory spelling in rules/field-types.md (DATA-E-03, +150, paid by DATA-D-14 + DATA-D-11) — the file whose job is field authoring contains zero Field.*; corpus is 92% factory.
3. required vs storage.notNull (DATA-F-02, +60) and 4 security-lint rows (DATA-F-04, +90), paid by DATA-D-15 + DATA-D-13.

Flight scope

IMPLEMENT (same-file, shrink-only): every DELETE / MERGE-INTO / REWRITE-AS-CONSTRUCT row in the findings table at HIGH or MED confidence; the three funded additions; INC-01/INC-02 corrections; os:check markers on the top-traffic examples with the seven would-fail blocks repaired (DATA-E-01/03/05); the hooks consolidation — delete rules/hooks.md, keep references/data-hooks.md as the single hooks file with its unique content preserved and paid by the data-hooks deletions; drop the deleted file's ceiling row.

ANCHOR RULINGS (this package is the anchor): objects, fields, relationships, hooks, seeds, datasources, permissions/RLS. Sibling flights (ui, api, platform, query, automation) delete their copies and point HERE — keep the paths rules/relationships.md, rules/datasources.md, rules/field-types.md, references/data-hooks.md stable. The inbound link rules/hooks.md:202 → objectstack-platform rules/plugin-hooks-events.md: the platform flight retires that file; re-point to objectstack-platform/references/plugin-hooks.md. DATA-C-03 (test harness) → delete + one-line pointer to objectstack-platform. DATA-C-01 → delete (query keeps its copy).

DEFER (pending #14296): DATA-I-04 (new rules/security.md), DATA-H-01 (evals ceiling / stub), DATA-F-07 (defineHook() prescription — maintainer call; record as a follow-up in the PR body), DATA-F-05 retire-surface candidates → shrink to one schema-pointing row, do not delete.

Flight constraints (binding)

• ONE draft PR, first line Fixes #<this card>; governed (skills/** md) ⇒ stays draft; review requests are the seat's step.
• Token ratchet: no ratcheted file may grow; additions paid by deletions in the SAME file; ⛔ re-wrap is not payment; ⛔ no ceiling raise; ⛔ no new files; scripts/check-skills-token-ratchet.mjs touched ONLY to drop the row of a file this PR deletes.
• ⛔ Never edit another package's files: MERGE-INTO another package = DELETE here + one-line pointer; a missing/weaker anchor goes in the PR body under "follow-up for ".
• Generated files untouched (references/_index.md); frontmatter edits ⇒ regenerate skills/README.md (pnpm --filter @objectstack/spec gen:skill-docs).
• Live surface with zero measured usage ⇒ one row pointing at its schema, never delete the last mention; retired/tombstoned surface ⇒ delete.
• A false claim matching a spec .describe() string ⇒ file the spec-side twin as an out-of-scope 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, plus node scripts/pm/dispatch-gates.mjs --commands <changed paths>; record the head sha.
• PR body: per-item 落点 | before | after list keyed by finding id; per-file token delta; needs:contract-review on both carriers if any operator/contract-semantics claim changes (INC-01 does).

Refs: #14292 · #14296 (batch-1 decisions) · #13658 (truth sweep).

[os-litant]

Claim: PM loop round 1
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14297-skills-data-optimization
Worktree: objectstack-issue-14297
Domain: domain:skills
File surface: skills/objectstack-data/** (delete rules/hooks.md; every other file shrink-only) + scripts/check-skills-token-ratchet.mjs ONLY to drop the deleted file's ceiling row; ⛔ no other package, ⛔ no .claude/** (stop on breach; explain in the report)
Container & model: M (judgment-heavy prose diet over 12 files), mode:subagent, model: opus — node scripts/pm/dispatch-gates.mjs --tier skills/objectstack-data/SKILL.md reports no path-derived mandate (floor sonnet · default opus · ceiling fable); the content limb applies (below), compensated by the seat's in-seat review at CONTRACT_REVIEW_TIER — the same posture the twelve truth-sweep flights landed under
Clause-②: yes — the flight corrects published contract-semantics claims (INC-01 required vs storage.notNull drift semantics; the sharingModel publish refusal) ⇒ needs:contract-review on both carriers at PR creation; in-seat review before any human review request
Serial constraints cleared: sibling flights #14298 (objectstack-ui) and #14299 (objectstack-platform) run in parallel on disjoint package directories; both point at this package's stable paths (rules/relationships.md, rules/datasources.md, rules/field-types.md, references/data-hooks.md); the platform flight retires objectstack-platform/rules/plugin-hooks-events.md — this flight re-points the rules/hooks.md:202 link to objectstack-platform/references/plugin-hooks.md; the ratchet-map rows touched by the three flights are disjoint; skills/README.md is NOT regenerated in-flight (catalog regen is a seat follow-up) unless the drift gate demands it — dev verifies and reports which

Decision re-read: no maintainer ruling on this card (filed this round); binding inputs = the mandate 「审核所有的 skills，进行全面的优化。」, the standing value gate, the token-ratchet discipline, and the batch-1 decisions #14296 (items 1–2 deferred as listed on the card).

---

Generated by Claude Code

[os-litant]

Claim: dev seat (os-dev), issue 14297
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14297-skills-data-optimization (pushed empty at claim time; base 909a4417)
Worktree: ../objectstack-issue-14297

Premise check before any edit: git diff --stat a59f78d HEAD -- skills/objectstack-data is EMPTY at base 909a4417, and git diff --stat d63c8a25 HEAD -- skills/ is EMPTY too — so every span in the audit record resolves byte-identically at my base, and scripts/check-skills-token-ratchet.mjs is unchanged since the audit pin. PM mechanism assumption 1 verified; spans are trusted as written.

File surface I will touch, per the card's binding constraints: skills/objectstack-data/** (delete rules/hooks.md; every other file shrink-only, generated references/_index.md untouched) plus scripts/check-skills-token-ratchet.mjs ONLY to drop the ceiling row of the deleted file. No other package, no .claude/**. PR stays draft; needs:contract-review on both carriers.

The full audit record follows as the next comment.

---

Generated by Claude Code

[os-litant]

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

Posted by the dev seat at claim time, from the audit seat's scratchpad audit/objectstack-data/findings.md. One mechanical transformation from that source: the five HTML-tag-shaped fragments were respelled as placeholders ([SolutionBlueprint], [!-- os:check --], KEY, FIELD) because the GitHub body sanitizer eats such fragments — including inside backticks — and can eat from the first one to the end of the comment. Nothing else was changed.

---

objectstack-data — published-skill audit (read-only)

Pinned ref: objectstack origin/main = a59f78d. (Working checkout is at a39b02a,
the immediate parent; git diff --stat a39b02a a59f78d -- skills/ is EMPTY, so every
span below is byte-identical at the pinned ref.)

Measured baseline

node scripts/check-skills-token-ratchet.mjs, objectstack-data rows:

File	tokens	ceiling	headroom
SKILL.md	13,892	13,892	0
references/data-hooks.md	12,584	12,611	27
rules/relationships.md	3,773	3,778	5
rules/field-types.md	3,578	3,584	6
rules/indexing.md	3,240	3,241	1
rules/validation.md	3,109	3,109	0
rules/hooks.md	2,195	2,195	0
rules/lifecycle.md	1,590	1,590	0
rules/datasources.md	903	911	8
rules/naming.md	773	773	0
evals/README.md	143	143	0
references/_index.md	1,483	—	generated

Hand-authored total 45,780; with the generated _index.md 47,263.
Bundle total (whole shipped tree) 187,854 → this package is 25.2 % of the
published catalog, and the largest of the eleven by 29 % (next: objectstack-ui
36,542; objectstack-platform 24,281; objectstack-formula 6,304). Total headroom across all 10 ratcheted files: 47 tokens.

Real-usage corpus measured (examples/** + packages/apps/**): 26 *.object.ts
files carrying 31 ObjectSchema.create() calls, ~250 Field.* calls,
3 defineSeed modules, 2 *.hook.ts, 2 *.datasource.ts, 0
*.seed.ts.

---

Findings

A — Trigger precision

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-A-01	A	skills/objectstack-data/SKILL.md:3-16 (frontmatter description)	The description enumerates "objects, fields, field conditional rules, relationships, validations, indexes, lifecycle hooks, permissions, row-level security … seeds". It never says datasource, external / federated object, or connect to an existing database — yet rules/datasources.md (903 tok) is entirely that surface, and skills/objectstack-platform/SKILL.md:206 is the only other mention in the catalog (a bare key in a list of stack keys). Platform's own description sells "drivers, adapters", so an agent asked "point my app at an existing warehouse Postgres" routes to platform and finds one word.	ADD external / federated datasources (defineDatasource) to the description, paid by DATA-A-03 in the same file	+18 / −287	① yes — decides which package is loaded for a whole rule file ② no ③ n/a (frontmatter is prose by format)	HIGH	yes (frontmatter → regenerates skills/README.md)
DATA-A-02	A	SKILL.md:3-16 vs SKILL.md:858-971, SKILL.md:556-582, rules/lifecycle.md	Three further surfaces have no trigger word in the description: metadata protection (SKILL.md:858-971, 987 tok), object extension defineObjectExtension (SKILL.md:556-582), and lifecycle retention/TTL/rotation (rules/lifecycle.md, 1,590 tok, only reachable via the pointer at SKILL.md:317). 2,577 tokens of published text with no routing sentence.	REWRITE-AS-CONSTRUCT: fold lifecycle + protection into the description's noun list; see DATA-C-05 for protection's real home	+12	① yes ② partly — DATA-C-05 removes one of the three ③ n/a	HIGH	yes
DATA-A-03	A/D	SKILL.md:26-55	Three consecutive blocks restate the frontmatter the agent has already read: 26-32 intro prose ("Expert instructions for designing business data schemas…", 71 tok), 34-43 "Skill Boundaries" 4-row table, 45-55 "When to Use This Skill" 7 bullets. The description already carries "Use when the user is creating or modifying *.object.ts / *.seed.ts files, picking field types, modelling relationships, writing beforeInsert/afterUpdate hooks…" and "Do not use for querying data (see objectstack-query) or for plugin / kernel hooks (see objectstack-platform)". The only content not in the description is two extra routing rows (api, ui).	DELETE 26-32 and 45-55; keep the 4-row boundary table	−287	① no — the agent loaded the package, so "when to load it" changes nothing ② yes, deletion is the whole fix ③ the surviving table is the construct	HIGH	no

B — Structure / progressive disclosure

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-B-01	B	rules/hooks.md (whole file, 2,195 tok)	Line 3 states its own status: "This document is a reference pointer. Complete documentation has been moved to the canonical hooks skill." Then it re-states the canonical file's content anyway. Measured overlap with references/data-hooks.md: hooks.md:29-44 ≈ data-hooks.md:93-136; hooks.md:46-50 is verbatim data-hooks.md:86-91; hooks.md:99-126 ≈ data-hooks.md:41-70; hooks.md:145-176 ≈ data-hooks.md:1050-1133; hooks.md:180-194 ≈ data-hooks.md:1136-1160. Genuinely unique content ≈ the condensed sandbox card at hooks.md:80-97 (314 tok) — itself a condensation of data-hooks.md:314-560.	MERGE-INTO references/data-hooks.md — keep hooks.md:80-97 as the SKILL.md-side card, delete the file	−1,880	① yes — two files disagreeing on which is canonical costs a re-read every session ② yes ③ the surviving card is the construct	HIGH	yes (removes a published file)
DATA-B-02	B	references/data-hooks.md:1-26	Two H1s in one file: # Data Lifecycle Hooks — Reference (L1) then # Writing Hooks — ObjectStack Data Lifecycle (L9), each with its own intro paragraph saying the same thing (L3-6 "Comprehensive guide to the 8 data lifecycle events, registration modes…" vs L11-13 "Expert instructions for third-party developers to write data lifecycle hooks"). L17-26 is a ## When to Use This Skill section — inside a reference, not a skill.	DELETE 9-26	−214	① no ② yes ③ n/a	HIGH	no
DATA-B-03	B	rules/hooks.md:99-126 ∥ references/data-hooks.md:41-70	The 8-event table AND the "why only 8 / bulk write hands hooks no row-scoping predicate / after* dispatch per matched row / no beforeFindOne / no *Many" callout appear in full in both files (466 tok and 525 tok).	DELETE one copy (subsumed by DATA-B-01)	−466	① no ② yes ③ n/a	HIGH	yes
DATA-B-04	B	rules/hooks.md:145-176 ∥ references/data-hooks.md:1050-1133	"Registration — three methods" (declarative / programmatic / hook files) stated twice, 144 tok and 566 tok.	DELETE data-hooks.md:1050-1133, keep the compact form	−566	① no ② yes ③ n/a	HIGH	yes
DATA-B-05	B	rules/hooks.md:180-194 ∥ references/data-hooks.md:1136-1160	✅DO / ❌DON'T lists, 5+5 vs 10+10 items, same content ("Use before* for validation, after* for side effects"; "Don't use object: '*' unless necessary"; "Don't assume ctx.session exists").	DELETE data-hooks.md:1136-1160	−418	① no — generic advice ② yes ③ n/a	HIGH	yes
DATA-B-06	B	SKILL.md:308-319	The single most decision-changing construct in the entry file — the 8-row pointer table telling the agent which rules file answers its question — sits at line 312 of 1,216, behind 300 lines of Core Concepts, an 88-line searchableFields essay and Field Groups. An agent that stops reading early gets tables it could have derived from the Zod source and never learns rules/ exists.	MOVE-TO immediately after the boundary table (SKILL.md:43)	0	① yes — changes what gets read ② no ③ the table already is the construct	MED	yes (reorders published file)
DATA-B-07	B	7 sites	"There is no unique validation type — use a unique index" is stated at SKILL.md:472-475, rules/validation.md:18-21, rules/validation.md:75-89, rules/validation.md:410-418, rules/validation.md:428, rules/indexing.md:45-79, rules/relationships.md:126-128. Seven statements of one rule.	DELETE the four restatements in validation.md (410-418, 428) and relationships.md:126-128's prose half; keep SKILL.md:472-475 + the indexing.md scope table	−150	① no after the first ② yes ③ n/a	HIGH	no

C — Cross-package redundancy

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-C-01	C	SKILL.md:111-190 (1,080 tok) ∥ skills/objectstack-query/SKILL.md:496-539	Both carry the "⛔ Searching by a related record's title — mirror the value, always" prescription, both open with the identical framing ("search scans the queried object's own columns. A dotted path is never a search target"), and both quote the same 400 error text verbatim (SKILL.md:180-186 = objectstack-query/SKILL.md:503-507, word-for-word). The two packages are loaded together for any list-search task. Data additionally quotes a second, 12-line os validate error verbatim at SKILL.md:165-176 — text the CLI prints unprompted.	DELETE SKILL.md:162-189 (both verbatim error transcripts + the "grep either back to here" framing). ANCHOR: data keeps the modelling half (the mirror field + the two-hook maintenance table at SKILL.md:151-160) — objectstack-query already routes to it by name ("Modelling side of this … objectstack-data → Search Fields")	−343	① no — the errors are printed by the tool at the moment they matter ② yes ③ the maintenance table survives as the construct	HIGH	yes
DATA-C-02	C	rules/validation.md:26-36, 313-331 ∥ skills/objectstack-formula/SKILL.md:101-141, 199-213	validation.md:28-32 teaches "CEL uses ==, !=, &&, `		, !— not SQL's=, AND, IS NULL" and 313-331 is a ❌/✅ pair on exactly that. objectstack-formula carries the same in depth (isBlank(v)table at :199, thehas()vsnull discussion at :119-141, "ISBLANK()`" retirement at :213), and this package's own description says "CEL expressions in formulas / validations … load objectstack-formula alongside".	DELETE validation.md:26-32 prose + the SQL ❌/✅ pair; keep only the inversion rule (34-36), which is validation-specific and lives nowhere else	−190	① no ② yes ③ n/a — ANCHOR is objectstack-formula
DATA-C-03	C	references/data-hooks.md:1209-1275 ∥ objectstack-platform's declared scope	"Testing Hooks" ships a vitest unit test and a LiteKernel + ObjectQLPlugin + DriverPlugin + InMemoryDriver integration test. skills/objectstack-platform/SKILL.md frontmatter claims exactly this: "operations (CLI commands, migrations, deployment, test harnesses via LiteKernel)". The block also imports four packages (@objectstack/core, /objectql, /runtime, /driver-memory) the example gate cannot resolve.	DELETE (ANCHOR: objectstack-platform owns LiteKernel harnesses)	−432	① no — a metadata-authoring agent does not stand up a kernel ② yes ③ n/a	HIGH	yes
DATA-C-04	C	25 tombstone sites; largest at SKILL.md:493-497, 742-746, 747-752, 797-803, 982-987, rules/indexing.md:26-32, 236-240, rules/field-types.md:266-267, rules/datasources.md:41-44	Narration of keys that no longer exist: type/partial retired, object-level rls removed, encryptionConfig pruned, tenancy.strategy retired, softDelete/versioning removed, referenceFilters removed, columnName removed, conditionalRequired removed, crypto.hash removed. Every one is enforced by a parse error that carries its own prescription (verified: packages/spec/src/data/field.zod.ts:864-865 ships the referenceFilters prescription as .guidance), and skills/objectstack-upgrade/SKILL.md exists solely for "when a parse or tsc error quotes a [REMOVED] prescription". A greenfield author never types these keys.	DELETE the standalone tombstone blocks (SKILL.md:742-746, 747-752 first sentence, 797-799, 982-987; indexing.md:236-240); keep only in-line ❌ markers where a wrong key sits next to the right one (indexing.md:141-149, field-types.md:266-267)	−450	① no for a new project; the compiler answers ② yes ③ the ❌/✅ pairs survive	MED (a maintainer may want tombstones for in-place upgrades)	yes (retired semantics)
DATA-C-05	C	SKILL.md:858-971 (987 tok)	The protection block is documented ONLY here, but SKILL.md:930-931 says it applies to "apps, views, dashboards, flows, agents, tools, skills, reports, email-templates" — every other domain. Grep of skills/** confirms zero prose coverage elsewhere (only references/_index.md generated pointers in ui/ai/automation/platform/i18n). Corpus usage: 0 occurrences of protection in any *.object.ts under examples/ or packages/apps/. Internally the section spends 3 near-identical 15-line examples (894-908, 912-926, 933-947) on one 4-value enum, plus an "Enforcement" subsection narrating Studio's ResourceEditPage banner (949-960) and a 5-bullet "Authoring guidance" (962-970).	REWRITE-AS-CONSTRUCT: keep the schema block (871-880) + the 4-row lock table (885-890) + ONE example; DELETE the other two examples, "Enforcement", and "Authoring guidance". Long-term: MOVE-TO a shared/platform anchor since it is not a data key.	−600	① the lock table changes a decision; the Studio-banner narration does not ② yes ③ table + one example is the construct	HIGH	yes (removes published capability text)

D — Density (true but decision-neutral)

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-D-01	D	references/data-hooks.md:1279-1322	"Performance Considerations" opens with an ASCII table of invented benchmarks — 0 hooks ~1ms / 5 hooks ~5ms / 20 hooks ~20ms — with no source, followed by 5 generic "Optimization Tips" ("Cache expensive lookups", "Batch operations") and an anti-pattern pair that restates async: true, already stated at data-hooks.md:199-221 and hooks.md:184.	DELETE	−337	① no ② yes ③ n/a	HIGH	no
DATA-D-02	D	references/data-hooks.md:1326-1381	"Advanced Topics": Dynamic Hook Registration (a forEach over config), Hook Composition (a for loop over an array of functions), Conditional Hook Execution (if (process.env.FEATURE_FLAG_ENTERPRISE !== 'true') return;). None teaches an ObjectStack fact; all three are plain JavaScript.	DELETE	−282	① no ② yes ③ n/a	HIGH	no
DATA-D-03	D	references/data-hooks.md:1385-1414, 1418-1444	"Troubleshooting" (4 generic issues: "Check object matches target object name") + "Summary" (6 ✅ bullets and 5 "Golden Rules" that repeat the DO list verbatim: "Use before* for validation, after* for side effects").	DELETE 1385-1414 and 1425-1444; keep the 2-line References at 1418-1422	−371	① no ② yes ③ n/a	HIGH	no
DATA-D-04	D	references/data-hooks.md:765-1046 (1,936 tok)	10 numbered "Common Patterns". Six carry no ObjectStack-specific decision: #1 defaults (if (!ctx.input.industry) ctx.input.industry='Other'), #2 validation (email.includes('@')), #4 enrichment (score += 50), #5 workflows (a console.log and a commented-out sendEmail), #7 external API (an entirely commented-out fetch), #9 conditional (a console.log under a condition:). Four DO carry facts: #3 (ctx.previous on delete + ctx.api.count), #6 (ctx.api.insert in-transaction), #8 ("There is NO updateMany — bulk updates use update(data, { where, multi: true })"), #10 (session.roles retired in 17.0.0, afterFind covers findOne).	DELETE patterns 1, 2, 4, 5, 7, 9; keep 3, 6, 8, 10	−1,000	① six of ten change nothing ② yes ③ the four survivors are the canonical examples	HIGH	yes
DATA-D-05	D	rules/indexing.md:81-106	"When to Add Indexes — ✅ Always / ⚠️ Consider / ❌ Avoid" is generic RDBMS 101 ("Low cardinality — Boolean fields", "High write volume — Every insert/update maintains indexes"). The one ObjectStack fact in it — foreign keys are NOT auto-indexed, declare them — is already stated at indexing.md:9-10 and again at indexing.md:330.	DELETE	−244	① no ② yes ③ n/a	HIGH	no
DATA-D-06	D	rules/indexing.md:161-234	"❌ Over-Indexing / ✅ Strategic / ❌ Wrong Order / ✅ Most Selective First" plus "Composite Index Strategy → Order Matters / Left-to-Right Rule / Selectivity Rule" — four ❌/✅ pairs and three headed subsections all teaching one textbook fact (composite indexes are used left-to-right).	REWRITE-AS-CONSTRUCT: one sentence + the ['status','created_at'] example already present at 109-121	−480	① one fact, stated seven ways ② yes ③ yes	HIGH	no
DATA-D-07	D	rules/indexing.md:277-340	"Performance Trade-offs" (✅/❌ benefit/cost bullets), a Table Size → Max Indexes table (< 1K rows → 2-3, > 1M rows → 8-12) with no source in the implementation, "Index Naming Convention", "Monitoring Index Usage" (raw pg_stat_user_indexes SQL and SHOW INDEX FROM), and a 10-item "Best Practices" list restating the file.	DELETE 277-298 and 312-340; keep the naming block 299-310 (name is a real declared key)	−430	① no — an authoring agent does not run pg_stat_user_indexes ② yes ③ n/a	HIGH	no
DATA-D-08	D	rules/relationships.md:398-413	"Best Practices" (7 items) + "Performance Considerations" (4 items) restating the file's own tables ("Use lookup by default", "deleteBehavior on master_detail — Always specify", "Index foreign keys", "Avoid deep hierarchies — tree relationships > 5 levels"). The last is an unsourced number.	DELETE	−264	① no ② yes ③ n/a	HIGH	no
DATA-D-09	D	rules/validation.md:420-438	"Best Practices" (10 items) + "Performance Considerations" (4 items). Item 7 restates DATA-B-07's rule a fifth time; item 9 ("Cross-field for comparisons — More efficient than script validation") and "Script validations are expensive" are unsourced performance claims about two rules that lower to the same CEL evaluator.	DELETE	−338	① no ② yes ③ n/a	HIGH	no
DATA-D-10	D	rules/validation.md:355-418	"Common Patterns": Prevent Backdating (= the script example at 39-52 with a renamed field), Require Approval for High Values (= the conditional example at 192-206 with a renamed field), Email Domain Whitelist + Phone Number Format (= the format example at 116-135 with different regexes), Composite Uniqueness (= 75-89). Five restatements of four examples already on the page.	DELETE	−313	① no ② yes ③ the earlier examples are the construct	HIGH	no
DATA-D-11	D	rules/naming.md:15-107	Six ❌/✅ pairs (78 lines) plus "Critical Rules" (5 restatements) plus "Rationale" ("snake_case for data: Database-friendly, SQL-compatible, cross-platform consistency"). The whole rule is the regex already in the table at naming.md:9-10 and again at SKILL.md:68 and again at SKILL.md:410-415.	DELETE 15-93 and 103-107; keep the table + Critical Rules	−566	① no — the regex is the rule; the pairs add nothing an LLM gets wrong ② yes ③ the table is the construct	MED (naming is a high-frequency error class; a maintainer may want one pair kept)	no
DATA-D-12	D	SKILL.md:540-555	"CRM Schema Blueprint (Production Pattern)" — a 6-row Pattern / Typical Location / Implementation Cue table whose rows point at fieldGroups, enable, indexes+validations, lookupFilters, hooks, state_machine — every one already covered above it or in rules/. Its two unique facts are tombstones ("no object-level workflows[]", "no separate stateMachines map").	DELETE	−358	① no ② yes ③ n/a	HIGH	no
DATA-D-13	D	SKILL.md:974-987	"Advanced Features Checklist" — 3 rows (tenancy, lifecycle, per-field trackHistory) each already documented in full above/in rules/lifecycle.md, plus a tombstone footnote (softDelete/versioning/partitioning/cdc/encryptionConfig).	DELETE	−234	① no ② yes ③ n/a	HIGH	no
DATA-D-14	D	rules/field-types.md:360-411	❌/✅ pairs for Wrong Type for Email, Uppercase Option Value, Missing Reference. The first two duplicate rules/naming.md:77-93 and the type table at field-types.md:17; the third duplicates the lint rule already listed at SKILL.md:1144 (relationship/missing-reference). Only the autonumber pair (412-427) teaches something the tables do not.	DELETE 360-411	−200	① no ② yes ③ the autonumber pair survives	MED	no
DATA-D-15	D	SKILL.md:1178-1181	objectstack lint --eval --generator ./gen.mjs — the live model-benchmarking door ("wire it to your agent / AIService.generateObject[SolutionBlueprint] … to benchmark a real model against the same rubric"). This is platform-benchmarking, not app authoring; a customer agent authoring account.object.ts never runs it.	DELETE	−110	① no ② yes ③ n/a	MED	no

E — Constructs & examples

The example gate is packages/spec/scripts/check-skill-examples.ts. Its header
(lines 17-23) states the rule: "this gate is opt-in: a self-contained,
should-compile block is marked by a [!-- os:check --] HTML comment on the line
directly above its fence", and it type-checks marked blocks with tsc --noEmit
against the built @objectstack/spec declarations, with module resolution wired
only from that package's own exports map.

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-E-01	E	package-wide	Measured fences: 151 ```typescript/```ts blocks, 5 os:check markers (SKILL.md:210, 250, 325, 513; references/data-hooks.md:512) = 3.3 %. Catalog comparison, same measurement: ui 11/33 (33 %), api 6/8, ai 4/6, i18n 4/10, formula 2/13, automation 1/11, platform 2/80, query 0/58. objectstack-data holds 41 % of the catalog's TypeScript and ~24 % of its markers.	ADD markers to the 5 blocks that would pass (see DATA-E-02), paid by DATA-D-01..03 in the same tree	+30	① yes — an unverified example is what a customer copies ② no ③ the marker IS the construct	HIGH	yes
DATA-E-02	E	17 unmarked self-contained fences	Blocks that carry their own import yet no marker. Would plausibly FAIL a type-check: rules/field-types.md:298-306 (an import followed by a bare { type: 'formula', … } in statement position — parsed as a block with labelled statements); rules/validation.md:39-52 (an import followed by validations: [ at top level — same defect); SKILL.md:1003-1030 (imports '../objects/status.object' / '../objects/category.object', unresolvable); SKILL.md:1096-1109 (uses Opportunity with no import → TS2304); references/data-hooks.md:1060-1070 (imports './objects/task.hook'); references/data-hooks.md:1213-1243 (imports vitest); references/data-hooks.md:1247-1275 (imports @objectstack/core, /objectql, /runtime, /driver-memory — outside the gate's paths map). Would plausibly PASS and should be marked: SKILL.md:564-575 (defineObjectExtension, exported at packages/spec/src/index.ts:181), SKILL.md:595-608 (definePermissionSet, index.ts:176), SKILL.md:894-908, 912-926, 933-947 (defineApp, index.ts:124).	REWRITE-AS-CONSTRUCT the two statement-position fragments (wrap in ObjectSchema.create({ fields: { … } })); mark the 5 passing blocks; the vitest/kernel blocks go under DATA-C-03	−20 net	① yes ② partly ③ yes	HIGH	no
DATA-E-03	E	rules/field-types.md (whole file, 3,578 tok)	The corpus authors fields with the Field.* factory — 250 calls measured (Field.text 61, Field.select 39, Field.lookup 22, Field.number 18, Field.currency 16, Field.date 14, Field.summary 11, Field.masterDetail 7 …) against 22 object-literal field declarations in examples/** + packages/apps/** (92 % factory; the literals are concentrated in the two *-zoo fixtures for types with no Field.* helper — tags, progress, repeater, vector). The factory is a first-class spec API with per-helper JSDoc (packages/spec/src/data/field.zod.ts:2120-2230, incl. the Field.lookup const-generic that defineSeed depends on). rules/field-types.md documents the literal spelling only — zero Field.* in the file. Meanwhile skills/objectstack-platform/SKILL.md:185-190, 635-655 teaches Field.text()/select()/date(). Two packages loaded together teach two spellings, and the data package teaches the one nobody uses.	ADD a Field.* column (or a 3-line preamble mapping literal ↔ factory) to the type tables, paid by DATA-D-14 + DATA-D-11	+150	① yes — it is the spelling of every real file ② no ③ table column	HIGH	yes
DATA-E-04	E	references/data-hooks.md:765-1046 vs rules/hooks.md:52-58	rules/hooks.md:52-58 says: "Prefer body for new code … handler (inline function) is deprecated", and SKILL.md:530-533 repeats it. All 10 canonical hook patterns (data-hooks.md:770, 794, 820, 848, 883, 911, 937, 968, 988, 1011) are written in the deprecated handler form. The single body example is data-hooks.md:512 (the one marked block).	REWRITE-AS-CONSTRUCT: after DATA-D-04 leaves four patterns, convert two of them to body	0	① yes — the package prescribes one form and demonstrates the other ten times ② no ③ yes	HIGH	yes
DATA-E-05	E	SKILL.md:325-379 (Quick-Start Template)	The canonical whole-object example carries name, label, sharingModel, enable, fields, validations, indexes. All 31 real objects additionally carry icon (31/31), pluralLabel (31/31) and description (31/31); 4 carry highlightFields. Conversely indexes appears in 1 of 26 corpus files. The template is the one marked, copy-pasteable object in the package and it diverges from every real one.	REWRITE-AS-CONSTRUCT: add icon / pluralLabel / description / highlightFields, drop the indexes block to the indexing rules file	+40 / −30	① yes — this is the block that gets copied ② no ③ yes	MED	yes

F — Gaps vs real usage

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-F-01	F	SKILL.md:64-84 (the object property tables)	sharingModel is used by 31/31 real objects (38 occurrences across 26 files) and is the single most frequent object key in the corpus after name/label/fields. It is absent from both property tables at SKILL.md:66-69 (required) and 71-84 (important optional). Its only prose in the entire catalog is a parenthetical inside the RLS notes at SKILL.md:738 and one unexplained line in skills/objectstack-upgrade/SKILL.md:634; it appears silently in 7 objectstack-data examples with no explanation. Implementation: packages/spec/src/data/object.zod.ts:2177 declares the canonical four (private / public_read / public_read_write / controlled_by_parent, "legacy aliases removed, ADR-0090 D4"), and packages/lint/src/authoring-rules.ts:1293-1296 records the maintainer ruling that made it effectively required: "an object publish with no authored sharingModel is refused with the 422 lint envelope (security-owd-unset); absence is not a decision." Rule id at packages/lint/src/validate-security-posture.ts:62.	ADD a sharingModel row to the required-properties table + a 4-row value table, paid by DATA-C-05 (−600) in the same file	+130	① yes — omitting it fails the publish ② no — nothing to delete instead ③ two table rows	HIGH	yes
DATA-F-02	F	rules/field-types.md (every required mention) + SKILL.md:383-403	storage.notNull (ADR-0113) is never mentioned in the package. packages/spec/src/data/field.zod.ts:918-925 declares it and states the split: "required governs what a WRITE must provide; storage governs what the COLUMN enforces", and the .describe() on required says "NOT a column constraint — the physical NOT NULL is a separate explicit opt-in (storage.notNull)". packages/drivers/driver-sql/src/schema-drift.ts:503, 795-798 confirms: "nullability drift reads THIS, not required". The package teaches required as the nullability dial 20+ times and never names the real one. See also the incidental falsehood below.	ADD one row to the object/field config guidance: required = write contract, storage: { notNull: true } = the DDL; paid by DATA-D-11	+60	① yes — decides whether a deployed column can be tightened ② no ③ one row	HIGH	yes
DATA-F-03	F	corpus keys with no coverage	Top-level object keys used in examples/**+packages/apps/** with zero prose in the package: icon (31 uses), pluralLabel (31 uses — one table cell at SKILL.md:76, no guidance), highlightFields (4 uses in 3 files; one parenthetical at rules/relationships.md:335, never listed as an authorable key), stageField (2), externalSharingModel (2), userActions (1), publicSharing (1), activityMilestones (1). highlightFields is load-bearing — relationships.md:334-335 says related-list columns "auto-derive from the child object's highlightFields" without ever saying how to declare it.	ADD highlightFields + icon rows to the optional-properties table; paid by DATA-D-12	+45	① yes for highlightFields (drives lookup pickers and related lists) ② no ③ table rows	HIGH	yes
DATA-F-04	F	SKILL.md:1142-1152	The lint table lists 9 data-model rules and 0 security rules, while the frontmatter claims the package covers "permissions, row-level security". packages/lint/src/validate-security-posture.ts:9-22 ships a 12-rule security-domain publish linter — security-owd-unset, security-owd-alias, security-external-wider-than-internal, security-wildcard-vama, security-anchor-high-privilege, security-role-word, security-master-detail-ungranted, security-cbp-no-relation … — of which the error rules "mirror a runtime enforcement point". An agent reading only the data skill's lint table believes lint-clean = the 9 rows.	ADD 4 security rows (security-owd-unset, security-owd-alias, security-external-wider-than-internal, security-master-detail-ungranted), paid by DATA-D-15 + DATA-D-13	+90	① yes — three of the four are publish-blocking errors ② no ③ table rows	HIGH	yes
DATA-F-05	F	speculative surfaces	Documented surfaces with zero occurrences anywhere under examples/** or packages/apps/** (what was grepped: grep -ho KEY $(find examples packages/apps -name '*.object.ts'), plus grep -rn over *.ts for the factories): tenancy 0 (SKILL.md:784-803, ~200 tok); protection 0 (SKILL.md:858-971, 987 tok); requiredPermissions 0 (SKILL.md:766-782); maskingRule 0 (named at SKILL.md:751); readScope/writeScope 0 (SKILL.md:652-688, ~340 tok); defineObjectExtension 0 (SKILL.md:556-582); declared indexes 1 file of 26 (examples/app-todo/src/objects/task.object.ts:201) against rules/indexing.md = 3,240 tok; sandboxed hook body: 0 (references/data-hooks.md:314-560 = 3,303 tok + rules/hooks.md:52-97). Zero usage in examples is evidence, not proof — tenancy/readScope are SaaS-deployment surfaces a demo app would not exercise, and body is the platform's stated forward direction for AI/Studio-authored hooks.	RETIRE-SURFACE candidates under the startup-focus principle: readScope/writeScope (half of it is behind a paid plugin — SKILL.md:681-688), defineObjectExtension. Others: trim, don't remove.	−400 (if taken)	① for a startup-stage catalog, no ② yes ③ n/a	MED	yes (removes published capability text)
DATA-F-06	F	SKILL.md:1129 vs SKILL.md:1004 and the corpus	Best-practice row: *"One {object}.seed.ts file per object	Readability at scale"*. find . -name '*.seed.ts' -not -path '*/node_modules/*' returns 0 across the whole repo. All three real seed modules are single files — examples/app-crm/src/data/index.ts, examples/app-showcase/src/data/seed/index.ts, examples/app-todo/src/data/index.ts — which is what the package's own example comment says (SKILL.md:1004: // src/data/index.ts). The frontmatter also fires on "*.seed.ts files". Two spellings of the convention inside one file, and the repo follows neither the row nor the trigger.	DELETE the 1129 row; ALIGN the frontmatter trigger to src/data/*.ts	−25	① yes — it prescribes a layout nothing uses ② yes ③ n/a	HIGH
DATA-F-07	F	rules/hooks.md:46-50 + references/data-hooks.md:86-91	"Prefer defineHook() over a bare : Hook literal" is stated twice. grep -rln defineHook --include='*.ts' outside packages/spec and packages/cli returns 0 files; both real hooks use the bare literal (examples/app-crm/src/hooks/opportunity.hook.ts:10 export const OpportunityStageHook: Hook = {; examples/app-todo/src/objects/task.hook.ts:51 const taskHook: Hook = {). The prescription is sound (import-time validation) but the platform's own examples contradict it — worth a maintainer decision, not a doc edit.	(no doc change) — flag for maintainer: either fix the two examples or drop the duplicate statement	0	① yes ② n/a ③ n/a	HIGH	yes

G — Consistency

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-G-01	G	SKILL.md:1135, 1171, 1174, 1178 vs SKILL.md:1195-1205, references/data-hooks.md:333, 467	Two spellings of the same binary in one package: objectstack lint/build (6 sites) and os lint/validate/build/migrate (13 sites) — including 60 lines apart inside SKILL.md (objectstack lint at 1135, os lint at 1202). packages/cli/package.json:17-20 declares both bins but :39 sets "bin": "os" as the oclif name, and the catalog runs 92 os * to 16 objectstack *.	REWRITE-AS-CONSTRUCT: normalise to os	−20	① marginal, but an agent copying two spellings into one README is noise ② no ③ n/a	HIGH	no
DATA-G-02	G	rules/relationships.md:265-286 vs the rest of the package	One fence uses the Field.* factory (Field.masterDetail('invoice', …), Field.lookup('product', …), Field.currency({ expression: … })); rules/lifecycle.md:34-91 uses Field.text({}) / Field.select([...], {}); SKILL.md:259-273 uses Field.select/datetime/currency. Every other fence in the package (≈140) uses object literals. Three spellings of field authoring inside one package.	Fold into DATA-E-03 (pick one spelling, state the other once)	0	① yes ② no ③ n/a	HIGH	no
DATA-G-03	G	rules/hooks.md:11, 200-208	Cross-package links are spelled ../../objectstack-data/references/data-hooks.md — a round trip out of the package and back in — for a file that is ../references/data-hooks.md. Both resolve (verified), but the file also links ../references/data-hooks.md#… at line 81, so both spellings appear in one file.	REWRITE (subsumed by DATA-B-01's merge)	0	① no ② yes ③ n/a	HIGH	no
DATA-G-04	G	package-wide grep for retired spellings	Grepped silently, ignored, dropped, legacy, deprecated, M9, v5.0, OLD, NEW, project-as-workspace, softDelete, versioning. No retired-spelling residue found: M9, v5.0, OLD, NEW return zero; every deprecated/legacy/removed hit is a deliberate tombstone (see DATA-C-04) or an accurate live statement (SKILL.md:501 bare unique: true on a declared index; rules/hooks.md:52 handler deprecated). "v5.0 vocabulary — runtime workspace is environment, not project" (skills/README.md) is honoured: SKILL.md:801-802 correctly says "environment/deployment choice".	(none — clean)	0	—	HIGH	no

H — Evals

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-H-01	H	evals/README.md:1-11 (143 tok)	The whole file: "⚠️ Not yet implemented — placeholder for future skill evals" followed by one sentence listing six topic keywords. No scenario, no expected output, no fail criterion, no rubric row. Comparison at the same measurement: objectstack-query/evals/README.md is 38 lines with 9 fully-specified evals each naming the expected construct AND the failure ("Expect having on the aggregation alias; fail on where, which filters input rows before the alias exists"); api 49 lines, platform 49, i18n 45, ai 44, automation 43 + an approvals/ fixture dir, ui 88 + a JSON fixture. objectstack-data is the largest package in the catalog (45,780 tok) with the smallest eval in the catalog (143 tok) — a 320:1 ratio against query's 42:1.	ADD 5 rubric rows testing the decisions this package exists to change — sharingModel present + correct value; master_detail vs lookup on a line item; script-validation inversion; unique index not a unique validation type; secret vs password for a credential — paid by DATA-D-01+D-02	+250	① yes — an eval that names no failure teaches a grader nothing ② no ③ rubric rows are the construct	HIGH	yes (raises a ceiling)
DATA-H-02	H	evals/README.md:6-7	The one candidate scenario that names a live decision — "secret vs password" — has no expected answer attached, so a grader reading it could as easily pass password for an API key. No row here would teach a grader to fail a correct answer, because no row states a criterion at all.	(folded into DATA-H-01)	0	—	HIGH	no

I — Size proportionality

id	dim	file:span	evidence	proposal	Δtok	value gate ①②③	conf	human-floor
DATA-I-01	I	package	45,780 hand-authored tokens = 25.2 % of the 187,854-token catalog, paid in every customer session forever. Distinct real-usage patterns found in F: 9 with ≥3 corpus usages (object+fields+sharingModel; field-type choice via Field.*; lookup; master_detail + inlineEdit; roll-up summary; relatedList; validations; field conditional rules; seeds via defineSeed+cel), 1 with 1-2 (indexes, hooks, datasources), 8 with 0. Measured deletions below total ≈10,300 tok (22.5 %) without touching any of the 9 live patterns.	(roll-up)	−10,300	① ② ③ per row	HIGH	yes
DATA-I-02	I	rules/indexing.md (3,240 tok)	7.1 % of the package for a surface with 2 occurrences in 1 of 26 corpus object files (examples/app-todo/src/objects/task.object.ts:201-204). The declaration surface is three keys — the file says so itself at indexing.md:13-25 — and the remaining 3,100 tokens are the ADR-0120 scope table (justified: unique: true on a declared index is deprecated and inverts meaning) plus ~1,400 tokens of generic RDBMS material (DATA-D-05/06/07).	Target ≈1,600 tok: three-key table + ADR-0120 scope table + the access-method/partial paragraph + one composite example	−1,600	① the scope table yes; the RDBMS 101 no ② yes ③ yes	HIGH	no
DATA-I-03	I	references/data-hooks.md + rules/hooks.md = 14,779 tok (32 % of the package)	One third of the largest package in the catalog is hooks, for a corpus with 2 *.hook.ts files, 0 sandboxed body hooks, and 0 defineHook() call sites. Within it: 3,303 tok on the sandbox ctx/capability contract (live and unique — keep), 1,936 on 10 patterns of which 6 teach nothing (DATA-D-04), 2,111 in a tail that is almost entirely generic (DATA-D-01/02/03 + B-05), and 2,195 in a second file that is 82 % duplicate (DATA-B-01).	Target ≈9,500 tok in ONE file	−5,300	① yes ② yes ③ yes	HIGH	yes (removes a file)
DATA-I-04	I	SKILL.md (13,892 tok, 1,216 lines, ceiling headroom 0)	The entry file is 30 % of the package and carries five sections that belong in rules/ or nowhere: searchableFields (88 lines), Security & Access Control (270 lines), Metadata Protection (113 lines), Seeds (141 lines), Linting (54 lines). Progressive disclosure exists (8 rules/ files) but the entry file did not shrink to match — it is the third-largest single file in the catalog (behind objectstack-ui/SKILL.md 25,441 and objectstack-pm-dispatch/SKILL.md 14,549) and sits exactly at its ceiling, so any true addition (DATA-F-01's sharingModel row) must be paid for inside it.	MOVE-TO rules/: the Security & Access Control block (585-854, measured 3,249 tok) → rules/security.md; keep a 15-line routing card	−3,100 from SKILL.md (net 0 to the package, but restores the headroom DATA-F-01/F-04 need)	① yes — it is what unblocks the two must-add rows ② no ③ n/a	MED	yes (restructures files, changes ceilings)

---

Incidental falsehoods (by-product of this audit, not the job)

INC-01 — SKILL.md:393, schema-evolution table.
Row: | `required: true → false` (relax `NOT NULL`) | dev auto-heals (`autoMigrate:'safe'`); otherwise `os migrate apply` |.
Two problems against the implementation:

1. required is not the NOT NULL dial. packages/drivers/driver-sql/src/schema-drift.ts:795-798:
   "nullability (ADR-0113: compared against storage.notNull, the explicit physical
   constraint — NOT against required, which is the write-time contract and implies
   nothing about the column)", with const expectNullable = field.storage?.notNull !== true;.
   packages/spec/src/data/field.zod.ts required.describe() says the same:
   "NOT a column constraint — the physical NOT NULL is a separate explicit opt-in
   (storage.notNull)".
2. It is not auto-healed. The drift the planner emits for a NOT NULL column with no
   declared storage constraint is kind: 'nullability_mismatch', severity: 'warning',
   category: 'needs_confirm', op: { type: 'relax_not_null' }
   (schema-drift.ts:815-826), and the comment above it says explicitly:
   "That surprise is worth a human decision; never auto-applied (a safe
   categorisation would let dev auto-reconcile strip a protection someone added)."
   autoMigrate: 'safe' applies only category: 'safe' ops
   (packages/cli/src/utils/schema-migrate.ts:491).
   Confidence HIGH. Human-floor: yes (touches migration semantics).

INC-02 — SKILL.md:398-401, the "tell-tale".
"/meta reports a field optional but a write still 400s "FIELD is required" — that
is a stale NOT NULL column (physical drift), not a validator bug." Per
schema-drift.ts:800-802, in exactly that configuration "an omitting write reaches the
DB and dies as a raw driver error instead of a clean validation 400" — so the symptom a
stale NOT NULL column produces is a driver error, not a "FIELD is required" 400.
Confidence MED (the message string was not traced to its producer). Human-floor: yes.

Both were found while measuring DATA-F-02; neither was re-derived by re-verifying claims
one by one.

---

Roll-up of proposed token deltas

Bucket	Δ
Hooks (B-01, B-03, B-04, B-05, C-03, D-01, D-02, D-03, D-04, B-02)	−5,300
Indexing (D-05, D-06, D-07)	−1,154
Validation + naming + relationships + field-types (B-07, C-02, D-08, D-09, D-10, D-11, D-14)	−2,111
SKILL.md (A-03, C-01, C-04, C-05, D-12, D-13, D-15, F-06)	−2,417
Additions that pay for themselves (A-01/02, E-01/03/05, F-01/02/03/04, H-01)	+825
Net	≈ −10,150 (−22.2 % of hand-authored)

---

Generated by Claude Code