skills(i18n): validation messages are a translation group since #14253 — correct the claim and its neighbours (net +7 tokens)

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14377

skills/objectstack-i18n/SKILL.md taught that validation messages are not a translation group. That was true when written and is false on main today: #14253 (landed as b2eab95, "i18n: give bulk-action defs, custom validation messages and dataset labels a bundle key") added three key families at once, and the published catalog ships to codebases this repo cannot see. This corrects the claim and the two neighbouring inventories it sits in.

A correction, not an expansion — maintainer ruling 2026-08-21, quoted verbatim and untranslated: 「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。」 Three new key families did not license three new sections: the diff is one added table row, one clause inside an existing row, one word in each of two existing prose lists, and one deleted blockquote. No new section. No ceiling change. The ratchet script is untouched, and the headroom #14574 left un-re-locked is not spent.

Per-claim: where, before, after

Line numbers are on origin/main at merge base 20b88391.

1. The claim itself — SKILL.md:172 (blockquote, 3 lines)

Before

validationMessages is not a translation group — it was removed in spec 17.0.0.
Author the message on the rule itself (object.validations[].message), which the
engine returns on every rejected write.

After — deleted; its substance moves into the new _validations row below, which is where a reader looking up an object sub-key will actually find it.

Why it is false.objects.OBJECT._validations.RULE.message is a declared group, object-scoped, with a reader:

• packages/spec/src/system/translation.zod.ts:427 — _validations: z.record(...) on ObjectTranslationDataSchema, sibling to _views / _actions / _tabs. message is the only declared key; label, description, condition and when each carry guidance rejecting them, so the surface is one key by measurement, not by omission.
• packages/spec/src/system/i18n-resolver.ts:1862 — objectValidationMessageKey(objectName, ruleName) returns objects.${objectName}._validations.${ruleName}.message.
• packages/objectql/src/validation/rule-validator.ts:2107 — the ObjectQL rule evaluator calls it on the write path and resolves through the existing ValidationMessageContext.translate hook (the i18nService channel that has localized built-in messages since Field validation messages are hardcoded English + API field name — penalty_amount must be ≥ 0 reaches end users verbatim #3957). Authored message on a miss, translation on a hit, no interpolation.

The schema's own note (translation.zod.ts:404-410) states the distinction the old blockquote lost: this is notvalidationMessages returning. That group was keyed by rule name at the top level, so it could not tell two objects' rules apart, and nothing read it — which is why 17.0.0 (#4667, ADR-0049) retired it. The new one is object-scoped and has a reader.

2. Object sub-key table — SKILL.md:163 (_views row)

Before	label, description, emptyState.title / emptyState.message
After	…unchanged…, bulkActions.{def_name} (label, confirmText, confirmLabel, params)

packages/spec/src/system/translation.zod.ts:281 — bulkActions is a record under _views.VIEW, keyed by BulkActionDefSchema.name. The four keys named are the ones the def actually declares; successMessage and description carry guidance rejecting them, and per-param help is spelled help, not helpText. The row names the declared keys so an author cannot reach for the two that parse-fail.

3. Top-level group inventory — SKILL.md:168-170

Before	apps, messages, globalActions, dashboards, pages, flows, settings, metadataForms, settingsCommon
After	same list with datasets inserted after dashboards

packages/spec/src/system/translation.zod.ts:770 — datasets is a top-level group on TranslationDataSchema, resolved by translateDataset. The inventory was incomplete by exactly one entry; it is corrected in place as a list item, not promoted to a section of its own.

4. Unknown-target inventory — SKILL.md:201

Before	…object, field, view, tab, action, param, section, app, nav item, dashboard, widget or flow screen…
After	…object, field, view, tab, action, param, section, validation rule, app, nav item, dashboard, widget or flow screen…

packages/lint/src/validate-translation-references.ts:629-634 registers _validations.RULE targets from objects[].validations[].name (the code names #14253 at the site). The lint side already reports an unknown validation-rule target; the enumeration simply had not been extended. One word.

premise_false — claims checked and left alone because they are true on main

• SKILL.md:387-391, the --objects-only divergence — true as written. --objects-only is on by default and --metadata-forms is orthogonal to it; packages/cli/test/i18n-extract-emitted-files.test.ts pins all three directions, including the regression where metadata-forms emission was briefly derived from --objects-only. Untouched — this is what 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) #14574 taught and it stays consistent.
• SKILL.md:380-386, the generated-bundle file shape — true as written. packages/platform-objects/src/apps/translations/ holds LOCALE.objects.generated.ts, LOCALE.metadata-forms.generated.ts and LOCALE.source-hashes.generated.ts exactly as described. Untouched.
• SKILL.md:407-436, the withSourceFallback wrapping and the kernel:ready load — true as written, and re-verified against packages/platform-objects/src/apps/translations/index.ts. Untouched.
• SKILL.md:462-465, the retired o.* sibling-key redirect list — correct to omitvalidationMessages. translation.zod.ts:470-480 defines LEGACY_OBJECT_FIRST_KEYS, and the schema comment at :532 states explicitly that validationMessages is "not a legacy object-first key". Adding it there would have been a factually wrong placement. Untouched.
• SKILL.md:281-285, the os i18n check surface list — the groups it enumerates are the groups the walker actually visits (packages/cli/src/utils/i18n-coverage.ts:185-218, COVERAGE_SOURCE). Untouched. See the note below.

Known gap, already filed — not folded into this PR

os i18n extract and os i18n check walk none of the three families #14253 added: packages/cli/src/utils/i18n-extract.ts emits no _validations, bulkActions or datasets path (its emitted roots are enumerated at :579, :649, :786-862, :928-1001, :1106-1173), and COVERAGE_SOURCE has no bucket for any of them. That is #14376 (open, domain:cli, pm:dispatched), already tracking exactly this. A caveat sentence about it was drafted into this file and then removed: it is the CLI's gap, #14376 is the place it is being fixed, and paying for it out of this file's budget would have made the net delta positive for a fact that belongs to another card.

Token budget — every added token paid by a deletion in the same file

node scripts/check-skills-token-ratchet.mjs, its own verdict line, before and after:

File	Before	After	Delta	Ceiling	Headroom after
skills/objectstack-i18n/SKILL.md	4678	4685	+7 (+0.15%)	6338 (unchanged)	1653
whole shipped bundle	157662	157669	+7	—	—

Lines: 496 to 493 (-3). Diff: 7 insertions, 10 deletions. The blockquote deletion is the payment — its substance is carried by the _validations row, so nothing true was dropped to make the budget.

• ⛔ The ceiling row was not touched; scripts/check-skills-token-ratchet.mjs is not in the diff.
• ⛔ The 1,660-token headroom left by 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) #14574's un-re-locked row is not spent — the re-lock remains that programme's step.
• ⛔ No re-wrap was used as payment; the only reflowed lines are the two prose lists that gained a word.

Gates

Head d2fa7a18fdb378722342f1cdba9e266f5bb797cd. Family re-derived after the last edit with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (change set: 1 path, skills/objectstack-i18n/SKILL.md; merge base 20b883918), run under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-14377. Every exit code captured before any pipe.

Gate	Exit	Its own verdict line
check-skills-token-ratchet	0	✓ 31 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted.
check:skill-identifier-liveness	0	OK — Leg 1: 465 citation(s) over 41 published file(s) checked against 93441 implementation word tokens (3 ledgered exemption(s)); Leg 2: 8 registered exhaustive section(s), 0 ledgered gap(s).
check:skill-docs	0	✅ Skill docs in sync
check:skill-compatibility	0	✓ 11 SKILL.md file(s) reconciled against 79 workspace packages
check:skill-frame-sync	0	✓ 4 copies of the decision frame are structurally isomorphic across 3 files
check:role-word	0	OK (self-test + scan; baseline unchanged, no --update)
check:corpus-claim-drift	0	OK
check:doc-authoring	0	✓ 14500 customer-facing string(s) across 710 spec sources clean
check:nul-bytes	0	OK (scanned 7995 text file(s); no raw ASCII control bytes)
check:agent-test-spelling	0	✓ 0 violations — 425 file(s)
check:pm-governed-merges	0	✓ self-test: 243 assertions
check-ci-filter-parity	0	OK: all 130 declared cross-package glob(s) (92 unique) are covered
check-cross-package-test-inputs / check:cross-package-test-inputs	0 / 0	OK: 25 package(s) read outside themselves, all declared
check-shard-attestation	0	✓ 2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s).

check:skill-docs was green without regenerating anything — the frontmatter did not change, so skills/README.md and content/docs/ai/skills-reference.mdx are not in this diff.

NOT MEASURED (prerequisite refusals, never content findings)

Gate	Exit	Prerequisite it named
check-test-completeness	3	PREREQUISITE NOT MET — it grades a saved turbo run test log and no log was named. Its own text says the family derives it with no argument and "the local reading for this gate is NOT MEASURED". CI tees the log and passes the path.
check:doc-formula-expressions	3	PREREQUISITE NOT MET — the workspace package @objectstack/lint is not built. (After building @objectstack/formula the refusal advanced to the next unbuilt package in the chain.)
check:skill-examples	1	PREREQUISITE NOT MET in substance — packages/client-react/dist holds no .d.ts declarations — the package is not built. Exit 1 here is the refusal branch, not a finding.

Declared narrowing, with the measurement behind it. The shared verify lock was continuously held by four sibling agents (two 9-minute queue timeouts at exit 99 and one foreground cap kill at 10 minutes, with three waiters ahead each time), so the remaining dependency builds those three gates need were not run locally. Rather than assert the narrowing, it is measured: for the two gates whose input is this file,

• check:skill-examples type-checks os:check-marked TypeScript blocks. os:check marker count in this file is 2 before and 2 after; fenced-block delimiter count is 24 before and 24 after; and git diff -U0 contains no added or removed line that is a fence or an os:check marker. The gate's own census reports 224 blocks on the skills + docs surface — this file's contribution to it is byte-identical to origin/main.
• check:doc-formula-expressions flags CEL field-rule expressions with rejected root identifiers across .claude, docs, skills, content. Every added line is reproduced in section 2-4 above; none contains expression syntax (no record., no current_user, no operator, no formula).

So neither gate's verdict can differ from main's on account of this diff. Both run on this PR in CI regardless, which is where the authoritative reading is.

Landing

skills/** is a governed surface (Prime Directive #14). This PR stays draft: no merge, no queue, no auto-merge, no ready flip. It is a PR visibly awaiting the maintainer's own merge. Labels: skip-changeset (verified against scripts/check-empty-changeset.mjs:357, which enumerates skills/ among the paths that release nothing) and needs:contract-review (the corrected rows state which keys the i18n resolver honours — a platform contract claim).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code