skills(i18n): optimization flight — drop the restated overview, the zero-usage layout and plugin bootstrap and the duplicate template; teach the generated-bundle shape and the --objects-only divergence (net −2,069 tokens)

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14315

Skills catalog optimization program #14292, maintainer mandate 2026-09-02: 「审核所有的 skills，进行全面的优化。」 Audit record posted verbatim on the card. Governed surface — this PR stays draft and awaits the maintainer's own merge.

Transliteration key. The GitHub body sanitizer eats angle-bracket-shaped fragments, so every one below is rewritten. [[X]] reads as "angle-bracketed X": [[os:check]] = the opt-in fence marker (an HTML comment whose text is os:check); [[locale]], [[o]], [[tab]], [[obj]], [[field]] = angle-bracketed placeholders; NonNullable[[TranslationData['objects']]] = NonNullable applied to TranslationData['objects']; Promise[[this]] = Promise of this.

---

Landing form — why Fixes and not Part of

The claim comment specified Part of #14315, for one stated reason: the deferred eval stub kept the card open. That reason is void. The maintainer ruled #142961A 2A 3A 4A, item 2 option A, which moved the stub from deferred to delivered — it is deleted in this PR (I18N-H-01 below). Every remaining DEFER item now has its own tracked home:

DEFER item	disposition
I18N-H-01 (eval stub)	delivered here under the 1A 2A 3A 4A ruling
I18N-C-03 (generated _index.md pointers)	filed as #14565, unassigned, spec lane
I18N-A-03 (routing bullet in skills/README.md)	recorded below as ruled, and filed as #14566 so it is tracked, not prose-only

Nothing is left on the card, so it closes on merge. That is also the state #14377 is waiting for — it carries Blocked-by: #14315 and becomes actionable once this lands. If the seat prefers the card kept open, this is a one-word edit before merge.

---

Per-item landing table

落点 = where it landed. Line numbers are the audit's (base 93940d49, byte-identical to the audit ref a59f78d for this package — verified, git diff --stat a59f78d origin/main -- skills/objectstack-i18n empty).

id	落点	before	after
I18N-B-03	SKILL.md:19-25	opening paragraph restating the frontmatter description	deleted; the H1 stays
I18N-B-04	:28-41	"When to Use This Skill", 7 bullets	deleted whole
I18N-A-01	:35-39	locale-formatting trigger (dates, numbers, currency)	deleted with B-04 — a whole-file grep for currency/timezone/format/Intl. matched only those 5 lines, so the trigger routed agents here for work the package teaches nowhere
I18N-B-01	:44-66	"Core Concepts" — 4 numbered items	deleted; each is restated in full by a later section
I18N-D-02	:99-148	3 prose "strategies" with ASCII trees	one 3-row table + a lead sentence
I18N-D-03	:143-148 and :526-530	the localesDir flat-file rule, stated twice	both deleted — localesDir has zero real usages and os serve never passes it
I18N-B-05	ordering	the layout prose sat between the config table and the object sub-key table	order is now config → canonical bundle → object sub-keys → naming → layout
I18N-F-03	:308-362	the translation metadata item, 570 tokens incl. a full worked example, 0 measured usages	one schema-pointing paragraph keeping the three rules that differ from a file bundle. Live surface, so the last mention stays
I18N-D-04	:356-362	the retired o.* dialect, stated twice	the Common Pitfalls copy survives — it alone carries the migration map
I18N-C-01	:506-530	hand-wired new ObjectKernel() + kernel.use(new I18nServicePlugin({localesDir, registerRoutes, basePath})) + bootstrap	one sentence on auto-registration + the 3 routes + a pointer to the platform anchor
I18N-E-01	:508-524	the one unmarked fence presenting as a complete program that would fail the example gate (used II18nService with the import in a different fence, TS2304)	dies with C-01
I18N-E-02	:513	kernel.use(...) un-awaited (hazard 1)	dies with C-01
I18N-D-01	:455-469	"AI-Powered Translation Suggestions" — vendor list + best-practice aside	deleted; also trimmed the 4th restatement of the same "contract-only" fact at the workflow's Translate step
I18N-D-05	:569-585	"CRM I18n Blueprint", a 4-row table restating three other sections	the two example paths it alone carried
I18N-D-07	:608-638	two pitfalls restating the naming callout and the option-value paragraph	deleted; the surviving paragraph is the stronger statement (it names the failure)
I18N-B-02	:646-688	"Quick-Start Template", a strict subset of the canonical example, as its own preamble conceded	deleted
I18N-C-02	:690-704	"Verify your work" restating objectstack-platform then pointing at it	2 commands + the pointer
I18N-D-06	:270-274	the validationMessages removal narrated with its history	one redirect line — see the collision note below
I18N-G-01	:300-302	lint unknown-target enumeration presented as exhaustive, _tabs missing	"tab" added; the lint walks objects.[[o]]._tabs.[[tab]] and reports translation-target-unknown
I18N-F-02 / I18N-G-02	:535-560, :546-547	extract → translate → check, documented as convergent; --objects-only absent from a 6-flag list	the flag is named, and a paragraph states the divergence (hazard 2)
I18N-F-01	:535-547 gap	the highest-count real pattern had no example — a reader following extract dead-ended at a file that cannot be registered	a canonical block: wrap the generated subtree, withSourceFallback, register or load at kernel:ready
I18N-F-04	:546 gap	--source-hashes had one word in a flag list	its purpose stated where the argument is passed
I18N-A-02	frontmatter description	triggered on none of what the body teaches for runtime authoring	triggers on the translation metadata item and on shipping a generated bundle
I18N-G-05	:247-248 vs the new block	(audit filed this repo-side)	the new block authors through defineTranslationBundle, so it no longer contradicts the file's own advice — see the gate note below
I18N-H-01	evals/README.md	45 lines / 410 tokens, an ASCII tree naming 6 files that do not exist, zero rubrics	file deleted, ceiling row dropped
I18N-B-06	evals/README.md:13-24	the fictional file tree	dies with H-01
I18N-G-03, I18N-G-04	—	"no action" in the audit	no action, except that G-03's verdict is falsified at this base — see the collision note

The two hazards

1. kernel.use is async.packages/core/src/kernel.ts:192 at this base reads async use(plugin: Plugin): Promise[[this]], and os serve's own call site awaits it (packages/cli/src/commands/serve.ts:2896). The skill's block did not. Fixed by deletion, not by adding await — the whole block was teaching a path nothing runs.
2. extract and check could not converge.os i18n extract's --objects-only is declared default: true, allowNo: true (packages/cli/src/commands/i18n/extract.ts:89-93), so extract writes only objects / globalActions; COVERAGE_SOURCE in packages/cli/src/utils/i18n-coverage.ts:192-220 also demands app, navigation, dashboard, widget, page and flow. Following the skill ended in coverage gaps the extract never scaffolded. Now stated, with --no-objects-only as the way out.

Anchor (I18N-C-01), cited by section

Plugin wiring / os serve auto-registration belongs to objectstack-platform → ## Runtime Boot Sequence, which carries the boot-tree line naming I18nServicePlugin. Verified by content on bothorigin/main and origin/claude/issue-14299-skills-platform-optimization (PR #14351, approved and may land first) — the section heading and the line are present on both, so the pointer survives whichever lands first. Cited by section, never by line number. Field and view labels stay in this package.

Why the generated-bundle block is NOT marked [[os:check]]

check:skill-examples compiles marked fences per surface, and the skills + docs surface is declared resolutionDir: SPEC_DIR, selfPackages: [SPEC_DIR] (packages/spec/scripts/check-skill-examples.ts, SURFACES) — it resolves @objectstack/spec and nothing else. @objectstack/spec does not depend on @objectstack/platform-objects (checked: not in its dependencies, no packages/spec/node_modules/@objectstack link), and the real shape needs withSourceFallback from @objectstack/platform-objects/apps. Marking the block would red it with TS2307 on correct code, so it is left unmarked and this paragraph is the reason.

It is verified anyway. A throwaway driver in packages/services/service-storage reproduced the block verbatim against the real exports and typechecked clean (tsc --noEmit, exit 0), twice — once for the original shape and again after the factory change below. Deleted in the same step both times; git status clean, confirmed by observing the files gone rather than by reading an exit code.

Symbols verified at source: withSourceFallback (packages/platform-objects/src/apps/translations/source-hash.ts:462, public via the apps barrel), the 4-argument order (translated, source, recorded, filledFrom), the generated module's NonNullable[[TranslationData['objects']]] export, and i18n.loadTranslations at kernel:ready (5 plugin call sites).

---

Token accounting — measured, not smoothed

Per the maintainer's 2026-08-21 ruling on published skills (「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。」「……不能为了一个小功能扩写很多。」), both readings:

The edited file, whole-file:

	before	after	ceiling	headroom
skills/objectstack-i18n/SKILL.md	6337 tok / 719 lines	4678 tok / 496 lines	6338 (unchanged)	1660
skills/objectstack-i18n/evals/README.md	410 tok / 45 lines	deleted	row dropped	—

The whole published package: 7296 → 5227 tokens shipped (−2069, −28.4%). Ratcheted total 6747 → 4678 (−2069, −30.7%). No ceiling was raised; no new file was added.

Per commit, against the audit's own estimates. Gaps are itemised per cluster, not netted away:

commit	finding ids	audit est	measured	gap
7164f69e	B-03, B-04 (incl. A-01), B-01	−633	−626	+7
1da918fb	D-02, D-03 (1st copy), B-05	−280 + part of −89	−234	see note
c97dafdd	F-03, D-04	−337	−395	−58
ff489c18	C-01, E-01, E-02, D-03 (2nd copy)	−150 + rest of −89	−105	see note
84a1928d	D-01, D-05, D-07	−448	−443	+5
74dd2f70	B-02, C-02, G-01, D-06	−357	−324	+33
36c7a16c	F-01, F-02, F-04, G-02	+205	+423	+218
4de6130e	A-02	+12	+33	+21
bc84ff2b	F-01 follow-up (G-05)	not in the audit	+12	+12
SKILL.md net		−1685	−1659	+26
d0dbe12c	H-01	−267 (as a rewrite)	−410 (as a deletion)	−143

Notes on the three gaps that are not rounding:

• 36c7a16c, +218 over. The audit costed F-01 at +180 for one block and F-04 at +0 ("fold one clause into F-01"). The block as written also carries the --objects-only divergence paragraph (audit: +25; actual closer to +70, because naming the six diverging key groups is what makes it actionable) and the --source-hashes rationale inline. I did not trim it to hit the number: it is the package's highest-count real pattern and its only convergence hazard. Flagging it rather than smoothing it.
• 1da918fb + ff489c18, +180 under (combined est −519, measured −339). Both are REWRITE-AS-CONSTRUCT items the audit costed as if only one line survived. The layout table keeps three rows and a lead sentence; the registration paragraph keeps the three route paths and the platform pointer. Both are smaller than a one-line estimate implies, and I judged the surviving content load-bearing.
• c97dafdd, −58 more than estimated. F-03's compression removed the whole worked example, not just :343-354.

On the dispatch's target of ≈ −1,418 on SKILL.md. That figure was the audit's −1,685 minus the deferred −267, but those are two different files: the audit's "Reducible total" gives −1,685 as SKILL.md's own net and −267 as evals/README.md's. The subtraction crossed files. Measured against the audit's actual per-file rows: SKILL.md −1,659 vs −1,685 estimated, and evals −410 vs −267 estimated (larger because the 1A 2A 3A 4A ruling deletes the stub where the audit had costed a rewrite).

---

Collision recorded and deliberately left alone: the object sub-key inventory, owned by #14377

While verifying premises at this base I measured that #14253 has landed, and it changes what is true in this file:

• ObjectTranslationDataSchema (packages/spec/src/system/translation.zod.ts:169) now declares a ninth object sub-key, _validations — objects.[[obj]]._validations.[[rule]].message, with a reader on the write path. packages/lint/src/validate-translation-references.ts:918-929 walks it.
• The same landing added objects.[[obj]]._views.[[view]].bulkActions.* and a top-level datasets group.

Consequences for this file, stated and left alone:

1. The D-06 blockquote is false today — validation rule messages are translatable now. My edit compressed that blockquote per the card's D-06 ruling and left its truth state exactly as it was. It did not pre-emptively repair it.
2. The object sub-key table still lists 8 sub-keys, and the lint-surface enumeration I extended with "tab" is still short by validation rules and datasets.
3. Audit finding I18N-G-03 is falsified at this base. It read "12 of 14 presented-as-exhaustive enumerations are COMPLETE" and passed the object sub-key table as COMPLETE. That was measured at a59f78d, before i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253. At 93940d49 it is not.

Why none of that is repaired here. The correction is owned by #14377, which instructs twice and explicitly (quoted verbatim, its emphasis):

⛔ Whoever takes #14315 must not pre-emptively fix line 270

⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

That card carries Blocked-by: #14315, and its triage ordering is "#14315 first, this correction second". Following the instruction is why the stale inventory ships one more round; recording the measurement here is so the next seat inherits it rather than re-deriving it.

---

premise_false

• I18N-G-03 — the audit's "object sub-key enumeration is COMPLETE" verdict, falsified at this base by i18n: three authored display surfaces have no bundle key at all (bulk-action defs, custom validation messages, dataset labels) #14253 (_validations, bulkActions, datasets). Recorded above; not acted on, per Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377.

Every other finding's premise held. Nothing was forced.

Follow-ups (not in this PR)

• follow-up for skills/README.md — I18N-A-03: the cross-skill routing list has no labels-vs-bundles decision point, so an agent authoring label / pluralLabel in *.object.ts or *.view.ts never learns this package exists (objectstack-ui contains zero occurrences of "translat"). Recorded here as the card ruled, and tracked as [finding] Nothing routes an agent authoring object/field/view labels to objectstack-i18n — the cross-skill routing list in skills/README.md has no labels-vs-bundles decision point #14566.
• follow-up for packages/spec — I18N-C-03: the generated references/_index.md routes 5 of 7 transitive pointers at schemas unreachable from a translation bundle. Tracked as [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565; sibling of [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 and [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466, same generator, adjacent defect (those are SKILL_MAP, this is the closure step).
• follow-up for skills/objectstack-i18n — the Published skill objectstack-i18n teaches that validation messages are not a translation group — false once #14253 lands #14377 correction described above.

---

Gates

Head sha bc84ff2b — the tree every reading below was taken from, and the tree pushed to this branch.

Card-named gates, exit codes captured before any pipe:

gate	result
node scripts/check-skills-token-ratchet.mjs	exit 0 — SKILL.md 4678 / 6338, headroom 1660
pnpm --filter @objectstack/spec check:skill-examples	exit 0 — 259 prose examples type-check across 3 surfaces (see the prerequisite note)
pnpm check:skill-compatibility	exit 0
pnpm check:skill-identifier-liveness	exit 0
pnpm --filter @objectstack/spec check:skill-docs	exit 0 — both artifacts in sync

check:skill-identifier-liveness was also checked before the first deletion: its BINDINGS table registers 8 exhaustive sections across 7 files, and none is in skills/objectstack-i18n (zero occurrences of i18n in the script), so no registered binding could be broken by a deletion here.

The full re-derived union was run, not just the card-named list.node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands was re-derived after the doc regeneration and after merging origin/main — the first derivation printed a STALE TREE warning naming 148 changed files, which is exactly the "well-formed answer about a tree nobody is on" failure, so it was re-run from a fresh tree. It prints 45 commands; all 45 were run on bc84ff2b.

Two results that are not green and are not red either:

• node scripts/check-test-completeness.mjs — exit 3, NOT MEASURED. The gate says so itself: "Nothing was measured … ⛔ It is NOT a finding". Reported as unmeasured, not as a pass.
• pnpm --filter @objectstack/spec check:skill-examples — first run exit 1 with PREREQUISITE NOT MET: packages/spec/dist/**/*.d.ts older than packages/spec/src, because merging origin/main moved spec's sources under an earlier build. Rebuilt the @objectstack/spec + @objectstack/client-react closures and re-ran on bc84ff2b: exit 0, "259 prose examples type-check across 3 surface(s) — every marked block parsed, so tsc ran the SEMANTIC pass on all of them". Reported as green only from the run that actually read fresh declarations.
• pnpm --filter @objectstack/spec check:docs — first run exit 1 with PREREQUISITE NOT MET (packages/spec/json-schema is a gitignored build artifact older than src). Satisfied it with gen:schema — run after the merge was committed, never in MERGE state — which left zero tracked drift, then re-ran: exit 0.

One gate caught a real defect in this PR's own addition, which is recorded rather than quietly fixed: pnpm check:doc-authoring went red on the new block, #2035 / ADR-0059, "Bare metadata-literal authoring found in docs/skills" — the block exported a bare : TranslationBundle literal, the one pattern that rule keeps out of the published catalog, and the same inconsistency the audit recorded from the other side as I18N-G-05. Commit bc84ff2b reauthors it through defineTranslationBundle; the gate is exit 0 on the final head, and the block was re-driven against the real exports after the change.

The merge of origin/main conflicted in both generated artifacts because sibling flights changed them. scripts/pm/os-regen-merge.sh correctly refused them (they are not os-regen driver paths), so the documented order was followed by hand: take main's side, commit the merge first, then regenerate with gen:skill-docs, then commit the regeneration separately. Asserted afterwards: every sibling package's row is byte-identical to origin/main's and both artifacts still carry all 17 rows — the staged diff is this card's description change and nothing else.

Labels

skip-changeset — this PR releases nothing. Checked against scripts/check-empty-changeset.mjs's own enumeration ("It releases nothing (.github/, .claude/, skills/, docs/, content/, examples/, tests-only, and the like)"): the five changed paths are three under skills/, one under content/, and one CI-internal gate script under scripts/ that publishes from no package.

needs:contract-review on this PR and on the card — the generated-bundle block, the --objects-only extract semantics and the un-awaited kernel.use hazard are contract claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code