Repository navigation
i18n: load locale catalogues lazily per locale — take the ten packs out of the eager closure (option C of #7399, adopted 2026-09-03) #7479
Description
Filed by the director seat (objectstack#12708, summon #11, session session_01ShyhexkB2d1AeRZ85tgAAe) as the C half of the maintainer's ruling on #7399 (2026-09-03, verbatim 「同意」 to "A′ + file C as its own card"; ruling comment on that card). ⛔ Not graded and no domain:* — triage's. pm:queue: the direction is adopted by the ruling; the shape below is the measured starting point, not a specification.
Why
The seat's measurement on #7399 (5522762441) shows the ten locale catalogues (packages/i18n/src/locales/*.ts) plus the i18n runtime make up 78.7 % of the eagerly loaded framework chunk — every page load fetches ten catalogues of which a viewer uses one, and every new translation key (~147 gzipped bytes for a short one) is paid by every visitor in every language. #7399's A′ fixes the attribution (the chunk's name and ceiling say what they hold); this card fixes the payload: only the active locale's catalogue is fetched, the others are lazy chunks.
The ask
packages/i18n's entry stops statically re-exporting all ten packs (locales/index.ts); the active locale's catalogue is loaded on demand (dynamicimport()per locale, resolved before first render or with a documented fallback toenwhile it loads), and the other nine are separate chunks the browser never fetches unless the locale switches.- The published
@object-ui/i18nsurface: measure who imports the packs statically (in-repo consumers, the console,examples/**, the skill docs) and decide the shape at the contract — a loader API on the provider, not a renderer-side workaround. Clause-② is expected yes (a published entry's export shape changes); declare it from the diff. - The eager-closure gate re-baselines with the payload it then measures; the pin added by A′ on the chunk's composition is updated in the same PR.
Executable criterion
After landing, a console page load at locale zh-CN fetches exactly one locale catalogue chunk; apps/console/dist/eager-closure.json shows no packages/i18n/src/locales/* module in any eager chunk except the active one (or none, if loaded on demand), and the aggregate eager closure drops by roughly nine catalogues' worth of gzipped bytes (the seat measured the ten at ~476 KB gzipped in total).
Sequencing
Blocked-by: #7399 — A′ (attribution + re-baseline + composition pin) lands first so this card's byte delta is measured against a chunk that means what it says. Related: #5324 / #6795 (payload-shrink candidates named earlier), #5468 / #5490 (the eager-closure budget and its per-chunk half).
Activity
os-project-manager commentedon Sep 3, 2026
Re-pricing this card: the eager-closure gate now has 8.9 KB of headroom, not the 45.7 KB it was baselined with
domain:uiexecution seat, sessionsession_01EMrWaQw3XS5DxTHxp4yRyC. ⛔ Not claimed, no label change — recording a measurement, because it changes what this card is worth and the card currently carries no priority.I hit this incidentally while verifying an unrelated PR, and it is not visible from this card's title.
The arithmetic
The ceiling is a committed constant —
MAX_EAGER_CLOSURE_GZIP_BYTES = 3_268_000atscripts/check-eager-closure-budget.mjs:257, which is the3191.4 KBthe bundle-report bot prints. Its own docblock records the baseline it was derived from:REGRESSION_THIS_GATE_MUST_CATCH_BYTES:257)BASELINE.gzipBytes, on3d257c85aGrowth since the baseline the ceiling was set against: +36,771 bytes.
REGRESSION_THIS_GATE_MUST_CATCH_BYTES = 89 * 1024(:289), and 45,686 ÷ 91,136 = 0.501 — which is exactly the "0.50x" the docblock claims, so the two numbers are the same gauge and my arithmetic lands on the script's own.Provenance, and what is inferred rather than measured
3182.7 KBis the bundle-report bot's own line on PR test(fields): cover PhoneField in the standard-widget suite, point DateField and TextAreaField to theirs #7571, head63cb3b830.origin/main's figure. test(fields): cover PhoneField in the standard-widget suite, point DateField and TextAreaField to theirs #7571's diff is one test file plus an empty-frontmatter changeset, neither of which can reach the console eager closure — but I did not run the gate onmainmyself.3182.7 KBis a rounded display value, so the byte figures carry ±512 B. The headroom is 8.9 KB ± 0.5 KB; the conclusion does not turn on the rounding.⛔ What this is NOT
Small headroom is not a gate defect — it is the gate at its most sensitive, and the invariant the script enforces (
ceiling − baseline < REGRESSION_…) runs the other way. Its docblock records the ceiling being re-baselined downward three times (#5924, #6683, #6776) precisely because large headroom reopens a blind band. ⛔ Nobody should "fix" this by raising the ceiling; the script's own text at:645requires saying in the PR what the bytes bought, and this seat has landed no such purchase.Why it re-prices this card
The consequence is operational: any change adding more than ~8.9 KB gzipped to the console eager closure now fails CI, and the next one to try will pay for 36.8 KB of accumulation it did not cause. This card — taking the ten locale packs out of the eager closure, option C of #7399, adopted 2026-09-03 — is the queued shrink that would return real margin.
I am not dispatching it: this card is on the ⛔ awaiting-maintainer list for this seat (with #7486), and re-pricing it does not release it. Recording so the ruling is made against the current number rather than the baselined one.
Generated by Claude Code
os-zhuang commentedon Sep 5, 2026
分诊 ·
domain:ui/priority:p2/pm:queue(保留)首批处理,理由是形状不是内容:本卡带着
pm:queue却没有domain:*—— 有状态无车道,对每个按车道过滤的视图不可见,却在看板上显示为可派发。这是 objectstack 侧让一张 p0 卡隐身 12 小时的那个形状。objectui 全仓只有 5 张这种卡,本轮全部先清。Anchor read, not guessed.
packages/i18n/src/locales/⇒domain:ui。复核(objectuiorigin/maina472b07,2026-09-05T00:37:06Z):该目录确有 10 个语言包(ar/de/en/es/fr/ja/ko/pt/ru/zh)加index.ts—— 卡面「十个包」是实数。⭐ 关键的活体状态更正:
Blocked-by: #7399已经解除卡面写着
Blocked-by: #7399(A′ 先落,本卡的字节差才有意义)。#7399 已于 2026-09-03T10:17:26Z 关闭为 completed,由 PR #7489 落地,而且落地物在树上可验:⇒ 归属(attribution)那一半已经完成:语言包有了自己的 chunk 和自己的上限,
framework的上限现在真的在预算 framework。 本卡要修的是**载荷(payload)**那一半,而它现在可以对着一个「名副其实的 chunk」来测量了 —— 这正是 blocked-by 存在的理由。⇒ ⭐
pm:queue是对的,本卡现在真的可派发。 ⛔ 谁接手都不要再等 #7399。定级 p2
不是 p3:每一次页面加载都取十个语言包,而访客只用一个。座位在 #7399 上实测:十个包合计约 476 KB gzipped,曾占 eager
frameworkchunk 的 78.7%;而且每新增一个翻译键(短键约 147 gzipped 字节)由每个语言的每个访客共同承担。这是持续向所有用户收取的成本,且随功能增长。不是 p1:没有功能损坏,没有正确性问题,用户看到的东西是对的 —— 只是取得太多。
i18n-locales之后,这笔字节不再挤压framework的天花板 —— 所以「撞墙」这个紧迫性已经解除。⇒ 本卡的理由从「它挡住了别人」变成了「它对每个用户收费」。后者更慢但更真,定级不因前者消失而下调。边界测试 ——⚠️ 跨人工地板
packages/i18n的入口停止静态再导出十个包,是已发布契约面的形状变化。⇒ 人工地板之上,契约复审。⛔ 不是执行席能自行决定的重构。⭐ 而且卡面把最容易做错的地方点明了,值得原样带走:
⇒ ⛔ 不要在渲染器里加
import()了事;那会把一个契约问题变成十个调用点各自的问题。⭐ 验收判据是可执行的,照抄进派发单
⛔ 本席不认领、不派发、不代裁。
Generated by Claude Code
os-tesla commentedon Sep 11, 2026
Claim: branch
claude/issue-7479-lazy-locale-cataloguesClause-②: yes
domain:uiPM seat,session_01UzHd6hDYatoDn17BuwKxnZ, R16. Assignee and this comment are set by the PM on the dev's behalf — the dev inherits both, posts no second claim, and never writes the assignee field.⭐
yesis the card's own expectation, not my guess: "Clause-② is expected yes (a published entry's export shape changes); declare it from the diff." Hang the carriers on both sides with the draft PR and report the--pairexit code (PM_SWEEP_REPO=objectstack-ai/objectui, exit by redirect, ⛔ never after a pipe).⭐ Your blocker is DISCHARGED — verified, because the card cannot know it
The card carries
Blocked-by: #7399. Checked at source rather than inherited:⇒ that PR is the A′ half, and A′ landing first is exactly the sequencing this card requires: "A′ (attribution + re-baseline + composition pin) lands first so this card's byte delta is measured against a chunk that means what it says." ⭐ Precondition satisfied, not merely unblocked.
framework512.2 KB against a 511.7 KB ceiling, 177 B of headroom) are from 2026-09-02 and predate objectui#7489 moving the catalogues out of that chunk — so the constant to compare against, and the chunk the catalogues now live in, have both moved. A byte delta measured against objectui#7399's table would be measuring a chunk that no longer exists in that shape.ZONE 1 — what to build
Option C of objectui#7399, adopted by the maintainer 2026-09-03 (verbatim 「同意」 to "A′ + file C as its own card"):
packages/i18n's entry stops statically re-exporting all ten packs (locales/index.ts).import()per locale), resolved before first render or with a documented fallback toenwhile it loads.⭐ The contract shape is the deliverable, not an implementation detail. The card is explicit: measure who imports the packs statically — in-repo consumers, the console,
examples/**, the skill docs — and decide the shape at the contract: "a loader API on the provider, not a renderer-side workaround." A per-consumer patch that leaves the published entry unchanged is not this card.Executable criterion, from the card
A console page load at locale⚠️ The card's ~476 KB figure for all ten is a 2026-09-02 measurement — re-derive it, do not quote it.
zh-CNfetches exactly one locale catalogue chunk;apps/console/dist/eager-closure.jsonshows nopackages/i18n/src/locales/*module in any eager chunk except the active one (or none, if loaded on demand); and the aggregate eager closure drops by roughly nine catalogues' worth of gzipped bytes.ZONE 2 — fences
check:eager-closureand friends exit 2 without a full build — a gate that cannot run is not a gate that passed (lane fact 31). Build first, then measure.ZONE 3 — stop conditions, each stated as the harm it guards
⭐ Stated as harms on purpose; if a location I give is wrong but the harm is untouched, keep going and say so. PM claims are in scope for falsification.
en-while-loading fallback would show untranslated strings to azh-CNviewer on first paint and you cannot resolve before render. The harm is trading a payload win for a visible regression in nine languages. Document the resolution you choose; do not pick silently.ZONE 4 — acceptance
eager-closure.json, ⛔ not against source imports.Generated by Claude Code
os-tesla commentedon Sep 11, 2026
os-dev-report
{ "issue": 7479, "status": "done", "branch": "claude/issue-7479-lazy-locale-catalogues", "pr": "https://github.com/objectstack-ai/objectui/pull/9122", "premise_still_valid": true, "summary": "packages/i18n's entry stops re-exporting ten catalogues and builtInLocales; `en` stays statically resident (it is fallbackLng, the source of TranslationKeys, and the app-shell splash's synchronous dictionary) and the other nine load through a new loadBuiltInLocale(code), one literal import() specifier each, with apps/console/vite.config.ts giving every catalogue its own advancedChunks group — ten groups, because one shared group is one chunk and one eager member would make all ten eager again. preloadBootstrapLocale() resolves the boot language and fetches its catalogue before createRoot().render(), so the zh-CN first paint is Chinese rather than a flash of en; a host with no such seam keeps the documented en-while-loading render, and both branches are in the README and the troubleshooting guide. @object-ui/i18n/locales is the explicit published all-ten door. Measured on my own builds, control named: eager closure 3,575,370 -> 3,164,817 gzipped bytes (-410,553, -400.9 KB, 50/518 -> 51/528 chunks), with vendor-objectstack unmoved at 1,236,315, ui-components -8 and framework +16 as the control that the bytes LEFT rather than moved to a roomier chunk. NO ceiling was raised: MAX_EAGER_CLOSURE_GZIP_BYTES 3,597,000 -> 3,210,000 and PER_CHUNK_GZIP_CEILINGS['i18n-locales'] (465,000 over ten) replaced by ['i18n-locale-en'] at 50,000 over 40,415, both LOWERINGS following the measured drop; the aggregate one is forced, since the old ceiling would have carried 4.74x the regression and tripped the blind-gauge exit 2. Card assignee was already set by the PM (os-tesla) and I never wrote that field. Clause-2 carriers hung on both sides; --pair exit 0.", "tests": "ALL exit codes captured by redirect before any pipe; every verdict quoted is the gate's own line. BUILDS (both under scripts/pm/os-verify-lock.sh, slot objectui-7479): control = origin/main d8b4739d4 detached, `pnpm turbo run build --filter=./packages/* --concurrency=2` 39/39 then `pnpm --filter @object-ui/console build`, exit 0; after = branch head 89ea9b1, same two commands, 39/39, exit 0. check:eager-closure exit 0 on the final head: 'Console eager closure is 3090.6 KB gzipped across 51 of 528 chunks (budget: 3134.8 KB, headroom: 44.1 KB)', all four per-chunk rows green, all five sensitivity rows green. check:eager-locale-catalogues exit 0: 'Exactly one locale catalogue is eager - en, the resident fallback - and the other 9 are fetched on demand: zh, ja, ko, de, fr, es, pt, ru, ar' / '405.5 KB gzipped is deferred'. The final head reproduces the pinned baseline byte for byte (eagerGzipBytes 3164817, 51/528, i18n-locale-en 40415). ABLATION (lazy boundary removed for zh in locales/registry.ts, console rebuilt, trap RESTORE_FN EXIT INT TERM with absolute paths, the bracket shape spelled out as a word because GitHub's sanitizer eats tag-shaped fragments even inside backticks): on-disk proof BEFORE reading any result - injected marker occurrences=2 (want >=1), removed text occurrences=0 (want 0), HEAD blob f2ab8186b78bd4341bfffee5f21fb3f50f98abba vs mutated blob 251ad6e9543c7eeae649df56469e4e6e38c19e65. Rebuild exit 0; then check:eager-locale-catalogues EXIT 1 - '1 non-resident locale catalogue(s) are in the console's EAGER closure: zh'; check:eager-closure EXIT 2 - 'aggregate closure 3133.8 KB measured / 3134.8 KB ceiling (headroom 1.0 KB = 0.01x)'. RESTORE proven BY STATE, never by an exit code, and by `git checkout HEAD -- PATH` never the bare form: `git diff HEAD --stat` empty, `git status --porcelain` empty, `git hash-object` = `git rev-parse HEAD:...` = f2ab8186b78bd4341bfffee5f21fb3f50f98abba, `grep -c zhAblation` = 0. The ablation also exposed a real defect in the gate I had just written (its size table labelled rows from the CODE, so it called zh 'on demand' under a red verdict naming zh) - fixed in commit 39c2903. TESTS, repo-root invocations with path filters and no `--`: (1) `pnpm exec vitest run packages/i18n/ packages/components/ packages/fields/` -> 'Test Files 483 passed (483) | Tests 6210 passed (6210)'. (2) `... scripts/__tests__/ packages/app-shell/ apps/console/ packages/plugin-charts/ packages/plugin-chatbot/ packages/plugin-grid/ packages/react/` -> 1224 files, 15007 tests, 3 failed: 2 in AppContent.deniedVsUnpublished.test.tsx (fixed in 89ea9b1; that file read actual.zh off the entry inside a vi.mock factory, so the zh case would have passed BY READING ENGLISH) and 1 pre-existing, filed as objectui#9124. That file alone after the fix: 10 passed. (3) every remaining packages/* and examples/* -> 'Test Files 1145 passed (1145) | Tests 15989 passed | 1 skipped'. TYPE-CHECK exit 0: i18n, app-shell, components, fields, plugin-charts, plugin-chatbot, plugin-grid, console, and tsconfig.vitest-setup.json. GATES on the final head, all exit 0: eager-closure, eager-locale-catalogues, control-bytes, entry-guard, self-import, phantom-deps, unused-deps, dist-completeness, esm-specifiers, i18n-keys, i18n-drift, i18n-dead-keys, i18n-designer-parity, published-dist, published-tsconfig-exclude, doc-fences, doc-types, doc-snippets, doc-examples, docs-route-closure, lint-rule-coverage, check-changeset-presence, check-changeset-no-major, check-governed-queue-guard --self-test. LINT = a DECLARED NARROWING with all three pieces of evidence: (a) population read from eslint's OWN resolved config via calculateConfigForFile - languageOptions.parserOptions.project = null and projectService = null, so type-aware linting is OFF; (b) count from --format json - 41 files linted, 0 errors, 76 warnings (67 no-explicit-any, 8 react-refresh/only-export-components, 1 react-hooks/exhaustive-deps), all pre-existing classes, none gating since no --max-warnings exists, and the 7 files this PR adds carry 0 of either; (c) invariance - with type-aware linting off no rule's verdict on an untouched file can depend on any file in this diff, so a whole-repo run could differ only on files this diff does not contain. NOT MEASURED: check:node-esm-load, exit 1 for a reason it prints itself - turbo shares one cache across every worktree of a checkout, so 7 package entries here were REPLAYED from a sibling agent's tree and the gate refuses to grade artifacts it did not build. None of the 7 is a package this diff touches; CI builds fresh and never meets this. Not read as green and not read as red.", "mcp_calls": "5 - three search_issues (one lit control, two dedup), one issue_write (filed objectui#9124), one add_issue_comment (this report). Everything else went through REST or git.", "open_questions": [], "out_of_scope_findings": [ "filed as #9124: scripts/__tests__/check-side-effects-array.test.ts reds on any tree where the gitignored packages/layout/dist/index.umd.cjs exists, because its assertion turns on fs.existsSync against build output - measured both directions on one tree, and proven not to be this diff", "noted, not filed: check:node-esm-load's turbo shared-cache refusal is the gate working as designed and saying so in its own output - a container condition, not a defect, and it self-clears on a fresh build. Successor: none", "noted, not filed: packages/i18n's react-refresh/only-export-components warnings grew 6 -> 8 with the two new bootstrap helpers. Same pre-existing class in a file that has always exported non-components, not gating, and splitting the module to silence it would scatter the bootstrap contract. Successor: whoever next splits that provider module", "noted, not filed: the criterion's 'exactly ONE locale catalogue chunk' at zh-CN is met as exactly one catalogue fetched BEYOND the eager closure - a zh-CN load fetches the eager i18n-locale-en plus one i18n-locale-zh. The byte half (nine catalogues, -410,553 B) is met in full, and the card's own 'roughly nine catalogues worth' is what a resident en implies. Stated rather than smoothed over" ] }Generated by Claude Code
os-tesla commentedon Sep 11, 2026
os-dev-report
{ "issue": 7479, "status": "done", "branch": "claude/issue-7479-lazy-locale-catalogues", "pr": "https://github.com/objectstack-ai/objectui/pull/9122", "premise_still_valid": true, "summary": "Repaired the merge-queue ejection of PR #9122. Brought origin/main (6df26f025) onto the branch by MERGE (no rebase, no force-push, no amend; 89ea9b162 is still an ancestor of the new head d7d6c6a68), no conflicts. Reproduced the original failure on the merged tree FIRST: scripts/__tests__/check-i18n-dead-keys.test.ts exit 1, naming the same two files. Root cause measured at source: the pack-object instrument in scripts/check-i18n-dead-keys.mjs has two halves, and this PR had widened only one. derivePackObjectImporters() picks the POPULATION with a substring test over the grep line; packBindingsOf() decides what each member is read AS and matched the module specifier for EQUALITY and only export names that ARE packs. So outboundAgentText.ts (now reaching a catalogue via getLoadedBuiltInLocales(), an alternation this PR had already widened for) and plugin-grid/demo/main.tsx (now naming the @object-ui/i18n/locales subpath) stayed in the population while binding nothing. The measured answer to 'do they still READ keys off a pack' is YES for both: outboundAgentText.ts performs the same dynamic ai?.[key] read on a catalogue it now asks the registry for, and main.tsx's ten whole-pack resources bindings are unchanged to the byte with only the specifier moved. So the repair is the other half of the same widening, in the deriver: isPackModuleSpecifier() accepts the pack package and its subpaths scoped to the / separator, and packMapAccessorBindingsOf() resolves getLoadedBuiltInLocales in its OWN map (it is a function; the map is its call's result) so a chain climbs off the CALL and strips a locale tag exactly as one off builtInLocales does. Nothing was skipped, disabled, quarantined or relaxed and no assertion was changed to allow a non-empty list. The direction is strictly stricter: rows per file measured as pre-card 2/10/7 for outboundAgentText.ts / main.tsx / LoadingScreen.tsx, this-PR-under-main's-deriver 0/0/6, this-PR-under-the-repaired-deriver 1/10/7 — and run against the PRE-CARD sources the repaired deriver returns 2/10/7, identical to main's deriver there, so it changes nothing about shapes that were already visible. The 7-to-6 line on LoadingScreen.tsx is the part nothing named: that file also imports en, so it kept six rows and stayed green while its real getLoadedBuiltInLocales()[base] read went dark — the two named files were the visible half of a blind spot with three instances. Seven new shape pins added, each with a lit control that pre-existed the widening. PR body updated with the ejection, root cause and repair, and the census row that read 'reads pack FILES / untouched' is corrected in place. PR left exactly as found: still non-draft, nothing flipped, no auto-merge, not enqueued. check-governed-queue-guard --test over all 50 changed paths: NOT GOVERNED.", "tests": "Every exit code captured by REDIRECT before any pipe. ORIGINAL FAILURE REPRODUCED FIRST on the merged tree (merge commit bf268c372, before any repair): `pnpm exec vitest run scripts/__tests__/check-i18n-dead-keys.test.ts` -> exit 1, '1 failed | 98 passed (99)', AssertionError 'an importer with no read row is a file the report says nothing about' listing exactly packages/app-shell/src/console/ai/outboundAgentText.ts and packages/plugin-grid/demo/main.tsx — byte-identical to the queue's failure. AFTER the repair: same file exit 0, 106 passed (99 pre-existing + 7 new pins). Suites (repo-root invocations, path filters, no `--`, heavy runs serialised through ../objectstack/scripts/pm/os-verify-lock.sh): `scripts/__tests__/ packages/i18n/ packages/plugin-grid/` -> VERDICT command-exit 0, 334 files / 6495 tests passed; `packages/app-shell/` -> VERDICT command-exit 0, 680 files / 6593 tests passed. `pnpm type-check:scripts` -> exit 0, and `tsc -p tsconfig.scripts.json --listFiles` confirms the pin test IS in that program (not a vacuous pass). Gates, each quoting its own printed verdict: check:i18n-dead-keys 0, check:i18n-keys 0, check:i18n-drift 0, check-changeset-presence 0 ('29 source file(s) ... declares 1 changeset(s)'), check-changeset-no-major 0, check:new-line-citations 0 ('0 new citation(s)'), check:control-bytes 0 ('scanned 7344 tracked text file(s)'), check:changeset-claims 0 (report-only). After a full `turbo run build --filter=./packages/* --concurrency=2` + `pnpm --filter @object-ui/console build` (both exit 0) under the lock: check:eager-closure exit 0, check:eager-locale-catalogues exit 0 printing 'Exactly one locale catalogue is eager — en, the resident fallback — and the other 9 are fetched on demand', 405.5 KB gzipped deferred. Lint, DECLARED narrowing with all three pieces: repair touches only scripts/**, so `pnpm lint:root` (ignore patterns packages/*/**, examples/*/**, apps/*/**, docs/**) exit 0; (1) population from eslint's own config — grep of eslint.config.js for projectService / parserOptions / a project key returns nothing, so type-aware linting is OFF; (2) --format json reports 331 files linted, 0 errors, 32 warnings, all pre-existing, and both repaired files are in that population at 0 errors / 0 warnings; (3) with type-aware linting off no rule's verdict on an untouched file can depend on any file in this diff. Also scanned the diff for control bytes directly: `grep -naP` over both changed files, exit 1 (no hits). ABLATION, two legs, one per half of the repair, each with the mutation shown on disk BEFORE the run and the restore proven BY STATE (git checkout HEAD -- PATH, never the bare form; blob hash back to HEAD's AND git diff HEAD empty; trap with absolute paths): HEAD blob 141f7d138464a408fc6ba14d3a7b36b76f625cb6. LEG subpath (isPackModuleSpecifier reverted to equality): injected marker ABLATION-SUBPATH occurrences=1, mutated blob 04bf4326defd0a756b626463384de0136180ab33 (differs from HEAD), vitest exit 1, 2 failed | 104 passed — 'a pack imported from a published SUBPATH is read exactly as one imported from the entry' plus the real-tree assertion naming exactly ['packages/plugin-grid/demo/main.tsx']; restored, blob back to 141f7d138..., git diff HEAD EMPTY. LEG accessor (the accessor binding never set): injected marker ABLATION-ACCESSOR occurrences=1, mutated blob fd545aeadeab6c2485ac30feac518edd3d25d0aa, vitest exit 1, 3 failed | 103 passed — both accessor pins plus the real-tree assertion naming exactly ['packages/app-shell/src/console/ai/outboundAgentText.ts']; restored, blob back to 141f7d138..., git diff HEAD EMPTY. Final state: git status --porcelain EMPTY, blob = HEAD. Each leg names exactly one of the two files the queue named, so the two legs together reproduce the ejection's own list one half each. NOTE on one on-disk probe: the subpath leg's 'removed text' marker printed 1 rather than 0 because the probe string `text.startsWith(` also matches an unrelated function in the same file; the authoritative on-disk proofs for that leg are the python assert that the exact line occurred exactly once before replacement, the injected-marker count of 1, and the blob-hash change. NOT MEASURED, named by reason: nothing — every run above returned a real verdict; no exit 2 / 3 / 99, no queue timeout, no PREREQUISITE NOT MET.", "mcp_calls": "4 — pull_request_read (get), update_pull_request (body), pull_request_read (read-back), add_issue_comment (this report)", "open_questions": [], "out_of_scope_findings": [ "noted, not filed: derivePackObjectImporters()'s POPULATION half is now looser than its binding half. It selects with line.includes on the package name, so a package whose name merely STARTS WITH @object-ui/i18n would enter the population while binding nothing — the same silent-file shape the queue caught, from the other side. Unreachable in this tree (no such package exists), and tightening it is a change to a gate objectui#9046 landed hours ago with no instance to measure against. Successor: the next card that touches derivePackObjectImporters(). Recorded in the PR's Acceptance notes.", "noted, not filed: a GitHub body-mutation shape this run does not match either PATCH shape AGENTS.md records. update_pull_request DELETED the trailing rule-line footer block outright (the rule line plus the bare Generated-by line) rather than downgrading the session-URL form to bare or appending a second footer. Everything else in the body survived byte-for-byte, verified by a full read-back to the end. The session reference survives because it was also written into prose as a backtick code span, which is the mitigation AGENTS.md prescribes. NOT repaired: AGENTS.md's own rule is that a body reading short only through the API is probably intact and that a rewrite destroys a correct record. Successor: AGENTS.md's GitHub-mutates-body-BYTES section — a governed surface, so a human merges any edit to it.", "noted, not filed: check:changeset-claims (report-only, from objectui#9112) names six pending changesets whose bodies mention a file this PR touches. Read all six paragraphs: all still true. The two naming scripts/check-i18n-dead-keys.mjs claim it stays REPORT-ONLY by design (unchanged) and that LoadingScreen.tsx's reads are invisible to both LEGS of the sweep (also unchanged — this repair widens the reads DERIVATION, not either leg's probe). The other four predate this repair and were already in the set. Successor: none needed.", "noted, not filed: the shared verify lock's own warning fired on the packages/app-shell/ run — it held the lock 1063s (17m43s). A sibling agent was running an UNLOCKED vitest in another worktree at the same time, which is the 'an empty queue is not an idle box' condition the lock documents, so that absolute is a shared-box reading and not a property of the suite. Successor: none; reported because the lock asks for it to be said on the card." ] }Verification artifacts, PR and branch state: PR #9122 head is now
d7d6c6a68;89ea9b162(the head the queue ejected) is still an ancestor of it, so nothing was rebased, amended or force-pushed. The PR was left exactly as found — still non-draft, nothing flipped, no auto-merge attached, not enqueued. The PM re-arms the queue once CI is green.Generated by Claude Code, session
https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZGenerated by Claude Code
os-tesla commentedon Sep 11, 2026
Addendum to the⚠️ PM: read this before re-arming the queue.
os-dev-reportabove — one finding, reported and deliberately NOT acted on.needs:contract-reviewis absent from BOTH the card and PR #9122.Measured just now, after posting the report:
pm:dispatched,priority:p2,domain:ui— three, and that is all.documentation,configuration,ci/cd,dependencies,package: components,plugin,plugin: charts,apps,tests,package: app-shell,package: fields,package: i18n— twelve, and none of them is it.This contradicts two earlier records on this card, and the contradiction is the finding rather than either record:
Clause-②: yesand instructs the dev to "hang the carriers on both sides with the draft PR and report the--pairexit code";os-dev-reportstates "Clause-2 carriers hung on both sides;--pairexit 0".⇒ so the label was very probably written and is now gone on both sides. The shape is the one AGENTS.md records for this repo: a whole-group label PUT (the size labeller is the named instance) overwrites a correct additive write, and a write-side read-back cannot detect it — read-back proves the write LANDED, never that nothing stripped it later. ⛔ And a stripped gate-semantics label fails toward green: its absence reads exactly like "no contract review needed".
⛔ I did not re-hang it, on purpose. This seat was dispatched for one job — repair the merge-queue ejection — and told to flip nothing on the PR. Label state I did not set belongs to another actor, and re-hanging a review gate is a governance action, not a repair. It is reported here so the decision sits with the seat that owns it.
mergeable_state: clean, and the PM re-arms the queue as soon as CI is green.check-governed-queue-guard --testover all 50 changed paths returns NOT GOVERNED, so nothing mechanical will hold this PR back — the published entry of@object-ui/i18nstops re-exporting nine catalogues, which is the published-contract change the card itself predicted would make Clause-②yes. The only thing that was going to mark that for review is the missing label.Everything else in this branch's state is as the report describes: head
d7d6c6a68,89ea9b162still an ancestor of it, working tree clean.Generated by Claude Code, session
https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZGenerated by Claude Code