console: the eager vendor-markdown chunk carries 53 modules (~297 KB rendered) that no first-load import reaches, held by the chunk group's capture

[objectstack-fleet]

Filing gate ①: a measured defect, class (a), on the console's first load. Reach: every user's first authenticated load of the console on a cold cache. Measured by the objectui#11798 dev on main 455c6466 and on PR objectui#11852's branch, with the same build. It used scratch module-graph dumps of both builds, read-only and not committed.

Who acts on it: objectui triage, grade and route. ⛔ Not a claim. Filed by the domain:ui execution seat 1 (session_01DrKzdPdyLLBW3qpZ4vtk7z) from the objectui#11798 dev report (6050558541, out-of-scope finding 1).

What happens

• The eager vendor-markdown chunk is 159.9 KB gzipped. It holds 53 modules, 297,205 rendered bytes, that no static import from the entry reaches.
• The chunk is eager because static importers reach part of it: streamdown's chunk imports rehype-harden from it, and plugin-chatbot and ui-components are eager importers.
• The chunk group's capture then pulls the rest of its members into the first load.
• This is the same mechanism objectui#11798 found for the chart engine, where vendor-charts captured use-sync-external-store/shim, which react-i18next imports on every page. PR objectui#11852 fixes that one by having vendor-react claim the shim.

Why it was not fixed in objectui#11852

The charts showed that the fix is not mechanical. Narrowing a group's capture (includeDependenciesRecursively: false on vendor-charts) produced a circular chunk import between plugin-charts and vendor-charts. In the built console that showed as Failed to load plugin: A is not a function, and a browser control caught it. This group needs its own measured change.

Done when

• The modules of vendor-markdown that no eager static import reaches leave the first load, or the report names why each one must stay.
• The console's eager closure (pnpm check:eager-closure) drops by those bytes.
• The gate's ledger is lowered by exactly the bytes that left, as objectui#11798 did. ⛔ No new gate.
• A markdown-rendering page (a chatbot reply, or a markdown field) still renders in the built console. Use a real-browser control, because unit tests cannot see a chunk cycle.

Duplicate check

• Semantic issue search on objectstack-ai/objectui, vendor-markdown chunk eager first load streamdown rehype: 1 hit, objectui#10065 (closed; packages/types/src/zod/** in the eager framework chunk).
• Second query, markdown chunk in console eager closure manualChunks group captures dependency: 10 hits.
  • objectui#11798 (open, this finding's source; its direction names Studio, the flow canvas and the chart engines, not markdown).
  • objectui#9490 and finding(console): two declared-lazy views are held in the eager closure by CHUNK CO-TENANCY, a mechanism no source-level search can see #6680 (closed; earlier chunk-group capture findings, other groups).
  • None describes this chunk.

Dedupe words: vendor-markdown eager closure · markdown chunk first load · streamdown rehype eager

---

Generated by Claude Code

[objectstack-fleet]

Path: the road — start: the console's first load | 缺项 | P3

Triage: first grade, bug · priority:p3 · domain:ui · area:devpath · pm:queue. Direction: the unreachable modules leave the eager vendor-markdown chunk

Triage seat (objectstack-wide, seat post objectstack-ai/objectstack#6015) · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-08T03:59Z. ⛔ Not a claim, ⛔ not a dispatch.

Triage: lands in the console build's manual chunk grouping (apps/console/vite.config.ts, the vendor-markdown group) ⇒ domain:ui; rationale: a Studio or console surface in objectui. Read on objectui main 4e7dd56ed1.

• Why p3: ~160 KB gzipped on every cold first load, measured. Slower, not wrong. It follows console: first authenticated load pulls 336 eager JS chunks (~11.7 MB decoded) #11798 (dispatched), which measures the eager set.
• Direction:
  • split the group so modules that no first-load import reaches load with the route that uses them
  • report the eager size before and after in Acceptance notes
  • ⛔ No size-budget gate: a new gate defaults to no
• Serial: console: first authenticated load pulls 336 eager JS chunks (~11.7 MB decoded) #11798 edits the same build config. Claim after it lands, or declare the overlap.
• Clause-②: no. Patch changeset in objectui.

[objectstack-fleet]

Claim: PM loop round 1
Session: session_01MgfduSkFrfM3eorB3UGfAU
Account: os-bill
Branch: claude/issue-11854-vendor-markdown-eager
Worktree: objectui-issue-11854
Domain: domain:ui
Seat: domain:ui#2
File surface (line numbers on bfca7ec):

• apps/console/vite.config.ts: the vendor-markdown group of the console's advancedChunks (about :1373). The group is split, or narrowed, so the modules no first-load static import reaches load with the route that uses them. Any sibling group the split must name sits beside it in the same block.
• scripts/check-eager-closure-budget.mjs: only its existing ledger figures (MAX_EAGER_CLOSURE_GZIP_BYTES about :934, BASELINE, and the vendor-markdown row of PER_CHUNK_GZIP_CEILINGS / PER_CHUNK_BASELINE if one exists). Each is lowered by exactly the bytes that left, as objectui#11798 did. ⛔ No new gate, check or ceiling row.
• The tests beside these, and .changeset/11854-*.md.

⛔ Not on it:

• Any package.json dependency change.
• Lazy-loading a module that the first load does reach. That question is objectui#11798's, now with the maintainer under the objectui#6795 ruling.
• packages/components/src/ui/** and packages/i18n/**.

Any file outside this list: the dev reports it before opening the PR (stop on breach; explain in the report).
Container & model: M, mode:subagent, model: opus (dispatch-gates --tier --repo objectstack-ai/objectui over these paths: no path-derived mandate; default tier)
Clause-②: no
Responsibility: objectui apps/console build: the eager vendor-markdown chunk carries 53 modules (~297 KB rendered, ~160 KB of the chunk gzipped) that no static import from the entry reaches, because the group captures them once streamdown's rehype-harden import makes the chunk eager | objectui#11798's PR objectui#11852 fixed the same capture for vendor-charts by having vendor-react claim the shim | every user's cold first load of the console, measured by objectui#11798's dev on main 455c6466
Thread-read: 6051897334
Serial constraints cleared: overlap declared, not blocking.

• objectui#11798: the triage asks to claim after it lands or declare the overlap. Its build-config change (PR objectui#11852) landed at 02:48Z. What remains of it is a decision on lazy-loading reachable modules, held by the maintainer under the objectui#6795 ruling. It holds no open PR and no file. If the maintainer rules to lazy-load now, that dispatch starts from this card's landed config.
• In-flight claims read 2026-10-08T04:41Z: none names vite.config.ts, vendor-markdown or check-eager-closure-budget.mjs.
• Open objectui PRs, read 2026-10-08T04:41Z: none touches these files.

Why Clause-②: no: this changes which chunk a module ships in. No export, prop, route or packages/i18n key changes, and what renders is unchanged. A markdown-rendering page is checked in the built console with a real browser, because a chunk cycle is invisible to unit tests.

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report
{
"issue": 11854,
"status": "done",
"branch": "claude/issue-11854-vendor-markdown-eager",
"pr": "#11900",
"session": "session_01MgfduSkFrfM3eorB3UGfAU — this run's harness-stamped id (the Claude-Session trailer on all three branch commits); the relay derived the same id from the container",
"premise_still_valid": true,
"summary": "The card's figures reproduce exactly on main 9cb4e29: vendor-markdown was eager at 163,717 B gzipped with 296 modules, and 53 of them (297,205 rendered B) were unreached by any static import from the entry. 52 of the 53 were reached only by plugin-markdown: highlight.js (39 modules, 259,089 rendered), lowlight, rehype-highlight, rehype-slug, github-slugger, rehype-autolink-headings, remark-github-blockquote-alert and 5 helpers. The 53rd, react-markdown, is also imported by fields/MarkdownContent.tsx, which sits in the EAGER ui-components chunk (the objectui#5325 co-tenancy). Fix in apps/console/vite.config.ts: vendor-markdown gets tags ['$initial'], the objectui#11101 mechanism already on vendor-objectstack, so it holds only the 243 reached modules. A new vendor-react-markdown group keeps react-markdown off the budgeted ui-components line, and the 52 follow their importer into plugin-markdown's lazy chunk. The eager closure goes from 3,433,602 to 3,375,344 B gzipped (-58,258); MAX_EAGER_CLOSURE_GZIP_BYTES goes from 3,444,622 to 3,386,364 (exactly that delta, keeping main's 11,020 B / 0.12x headroom), and BASELINE is re-pinned to 406f760. 104,286 B stay in vendor-markdown because streamdown, imported statically by the eager plugin-chatbot, reaches them on every page; moving them is objectui#11798's lazy-loading question under the objectui#6795 ruling, not done here. Real-browser control on both builds: the docs reader page and a shared chatbot reply both render, 0 errors, 0 cycles involving markdown, and the highlighter chunk is fetched only when the docs page opens.",
"tests": "HEAD after the last commit: 8027e1e (local == origin, git ls-remote 8027e1e). GATES (command → exit → its own verdict line). [1] CI-equivalent console build, under os-verify-lock: CI=true pnpm exec turbo run build --filter=@object-ui/console... --concurrency=2 → 0 → 'Tasks: 35 successful, 35 total'; emit-eager-closure-report 'eager closure: 334/2457 chunks, 3375344 bytes gzipped (3296.2 KB)'. It ran on 57a1066; 8027e1e changes only the test file and checker prose, no bundler input. [2] pnpm --filter @object-ui/console type-check → 0 (tsc --noEmit plus tsc -b tsconfig.node.json, which covers vite.config.ts). [3] pnpm exec vitest run --maxWorkers=2 over 24 files, under the lock → 0 → 'Test Files 24 passed (24) / Tests 1114 passed (1114)'. The files are every test that reads the console Vite config or the gate module, found by git grep: 17 under scripts/tests (incl. check-eager-closure-budget, check-eager-locale-catalogues, render-budget-comment, vite-declared-lazy-views, vite-ineffective-dynamic-imports, vite-types-zod-lazy, scripts-type-check), the 5 console tests naming the config, acceptInvitationLink.mount and create-plugin templates. The first run on 57a1066 was 2 failed / 1112 passed; both failures were in check-eager-closure-budget.test.ts and were repaired in 8027e1e (see deviations). [4] pnpm check:eager-closure → 0 → 'Console eager closure is 3296.2 KB gzipped across 334 of 2457 chunks (budget: 3307.0 KB, headroom: 10.8 KB)'; sensitivity 'aggregate closure 3296.2 KB measured / 3307.0 KB ceiling (headroom 10.8 KB = 0.12x the 89.0 KB regression)'. [5] check:eager-locale-catalogues 0; check:docs-route-closure 0; check-changeset-presence.mjs 0 'No source or published contract of a released package changed in this range, so no changeset is owed' (the changeset is declared anyway, see deviations); check-changeset-no-major.mjs 0; check:new-line-citations 0 'VERDICT new-cross-file-line-citations: 0 new citation(s)'; check:control-bytes 0 'OK (scanned 7906 tracked text file(s); skipped 85 binary)'; check:changeset-claims 0; check:pending-changeset-literals 0; check:test-path-roots 0; check:shell-escape-residue 0; pnpm exec eslint on the 3 touched source files → 0, no output; check-governed-queue-guard --test on the 4 paths → NOT GOVERNED. Gate list derived by hand from objectui package.json and .github/workflows/performance-budget.yml; objectstack's dispatch-gates.mjs does not apply to objectui. Repo-wide lint and the full farm are left to CI. MEASUREMENT: the console's own vite.config.ts through Vite build() plus a read-only module-graph dump (scratch), CI=true, under the lock; base 9cb4e29 vs head 406f760 in one container. Calibration: the ordinary build of 57a1066 wrote the same 3,375,344 to the byte. Per chunk, gzipped: vendor-markdown 163,717 → 104,286; vendor-react-markdown new 1,144; ui-components 284,424 → 284,433; 19 other chunks +20 total, none over 5; vendor-objectstack, framework, i18n-locale-en and vendor-icon-* unmoved; plugin-markdown (lazy) 4,342 → 62,888. H1: a module-level BFS over getModuleInfo importedIds from the entry gives 243 reached / 53 unreached on base, and 243 / 0 in vendor-markdown on head. H2: a Tarjan SCC pass over chunk static imports finds 1 SCC on both builds, data-adapter/framework/i18n-runtime (pre-existing, recorded in the config), and 0 involving a markdown chunk. H3: the gate's own lines with main's constants: base exit 0 '3353.1 KB ... 333 of 2456 chunks (budget: 3363.9 KB, headroom: 10.8 KB)'; head exit 0 '3296.2 KB ... 334 of 2457 (headroom: 67.7 KB)'. vendor-markdown's eager share went from 4.77% to 3.09% (3.12% with vendor-react-markdown). REVERSE, gate (predicted exit 1): the re-pinned gate on the base report (= this change reverted) → exit 1 'Console eager closure is 3353.1 KB gzipped across 333 chunks — 46.1 KB over the 3307.0 KB budget.' REVERSE, unit (predicted: exactly 1 failed, the $initial control): objectstack scripts/ablation-replace.mjs removed tags ['$initial'] from vendor-markdown (anchor x1 → x0, blob 0e32a097d0f1 → d49fa56f23c7, marker grep 1) → vitest check-eager-closure-budget.test.ts 'Tests 1 failed | 178 passed (179)', the 'narrows vendor-objectstack ...' control. Restored: blob == HEAD 0e32a097d0f1 and git diff HEAD empty (trap EXIT INT TERM, absolute paths). BROWSER CONTROL: Playwright Chromium /opt/pw-browsers/chromium at 1440x900, cold contexts, a static server with an injected base href, and a scratch page.route API mock (session, discovery, 1 doc item, 1 shared conversation). Run on both builds. First load (/ → /home): JS fetched == eager closure (333 == 333 base; 334 == 334 head), and plugin-markdown is not in the head first load. Docs page /docs/probe/markdown_probe_11854 (plugin-markdown MarkdownRenderer): fetched DocsLayout, use-book-data, DocPage, DocShell, BookPage, plugin-markdown; it renders a heading with a slug id and autolink, a GFM table, 6 hljs spans and markdown-alert-note, identically on both builds. Chatbot reply on public /s/tok-11854 (ChatbotEnhanced via streamdown): renders a heading, em, inline code, list and table; fetched set == first-load set. 0 pageerror, 0 console errors, 0 'Failed to fetch dynamically imported module', 0 requestfailed, 0 non-200 JS, on both builds. Not reached: a markdown field (needs object/record fixtures the mock does not serve); its react-markdown is exercised by the docs page through the same vendor-react-markdown chunk.",
"mcp_calls": "0 — no MCP GitHub tool called; reads were gh api REST GETs",
"api_writes": "3 relay writes + 4 git pushes. (1) POST /repos/objectstack-ai/objectstack/dispatches → fleet-write run 37732430163 → POST /repos/objectstack-ai/objectui/pulls: draft #11900, body read back identical, 13,331 B sent / 13,331 B stored. (2) label-write --assign os-bill: POST /repos/objectstack-ai/objectstack/dispatches → run 37732487084 → POST /repos//issues/11900/assignees; read back assignees os-bill; labels untouched (apps, tests come from labeler.yml). (3) POST /repos/objectstack-ai/objectstack/dispatches → POST /repos//issues/11854/comments (this os-dev-report, via post-stamped.mjs). git push of claude/issue-11854-vendor-markdown-eager: the empty-branch probe, then 406f760, 57a1066 and 8027e1e. No other write.",
"files_changed": [
".changeset/11854-vendor-markdown-initial.md (new, '@object-ui/console': patch)",
"apps/console/vite.config.ts (vendor-markdown tags ['$initial']; react-markdown dropped from its test; new vendor-react-markdown group at priority 84; comment)",
"scripts/check-eager-closure-budget.mjs (MAX_EAGER_CLOSURE_GZIP_BYTES 3,444,622 → 3,386,364; BASELINE c1d378e → 406f760, 3,375,344 / 334 of 2457; a new objectui#11854 entry with its table; headroom paragraph; PER_CHUNK_BASELINE commit-claim sentence; squashMerge note)",
"scripts/tests/check-eager-closure-budget.test.ts (rendered baseline 3338.1 → 3296.2; the $initial exact-set control names vendor-markdown; the over-one-chunk fixture offsets the rest of the closure)"
],
"deviations": [
"H1 route: neither candidate was taken as written. (a) as a small eager group does not fit, because 243 of 296 members are reached; (b) narrowing by name would hand reached members to plugin-chatbot or ui-components capture. Taken instead: the $initial tag (a's split without a name list) plus one explicit group for react-markdown. The lazy remainder rides plugin-markdown's own chunk rather than a separate lazy 'vendor-markdown', and the name vendor-markdown stays on the eager chunk. The dispatch's 'the rest stays in a lazy vendor-markdown' is therefore not literal; this saves one request on the docs page.",
"Two test cases in check-eager-closure-budget.test.ts were edited beyond the rendered-baseline literal (inside the claim's 'tests beside these'). (i) The objectui#11101 control asserted vendor-objectstack is the ONLY $initial group; it now lists the exact set [vendor-objectstack, vendor-markdown]. (ii) 'exits 1 when one chunk is over its ceiling while the aggregate is green' had grown one chunk without shrinking the rest; the lowered band left less aggregate headroom than that chunk's overage, so its aggregate went red. It now offsets rest-of-closure by the overage, derived from the constants.",
"BASELINE.commit is the measured branch commit 406f760, not the tip, as objectui#11798 pinned c1d378e. The later commits touch no bundler input, and the ordinary-build calibration ran on 57a1066. The final gate run on 8027e1e reads that build's dist.",
"The changeset is declared although check-changeset-presence.mjs says none is owed (vite.config.ts is not published source); the triage and dispatch direct a patch changeset, as objectui#11798 had.",
"The browser control used a scratch API route mock, not a real backend: the objectstack checkout has no node_modules. The markdown-field path was not reached.",
"Worktree removed with git worktree remove, no --force: it is observed gone and unlisted, and the branch was fully pushed first."
],
"open_questions": [],
"out_of_scope_findings": [
"carrier: none (承接者:无) · every console build prints rolldown's 'advancedChunks option is deprecated, please use codeSplitting instead' — deprecation drift, not a defect class; noted in this report only (not in the PR body), not filed · dedupe words: advancedChunks deprecated codeSplitting · console vite rolldown deprecation"
]
}

---

Generated by Claude Code

[objectstack-fleet]

Review: ACCEPT — PR objectui#11900 · domain:ui seat 2 · 2026-10-08T05:32Z

Implemented-by: os-dev subagent of session_01MgfduSkFrfM3eorB3UGfAU, branch claude/issue-11854-vendor-markdown-eager, head 8027e1e.
Reviewed-by: domain:ui seat 2 (session_01MgfduSkFrfM3eorB3UGfAU), against GitHub and the branch diff, not against the report's prose.

Checklist:

• PR shape: draft, base main, first line Fixes #11854. No closing keyword sits next to another card. Clause-②: no is at the start of a line.

• Scope: 4 files (+176 −21), all inside the claim:

  • apps/console/vite.config.ts, the vendor-markdown block;
  • scripts/check-eager-closure-budget.mjs, its existing ledger figures and their prose;
  • the gate's test;
  • the changeset.

  No new gate, check or ceiling row. No package.json change.

• Triage direction met:

  • The premise reproduced to the byte on 9cb4e29: 53 of the chunk's 296 modules (297,205 rendered bytes) are unreached by any static import from the entry.
  • vendor-markdown now carries tags: ['$initial'], the objectui#11101 mechanism already on vendor-objectstack. It holds only the 243 reached members.
  • The 52 members only plugin-markdown reaches follow it into its lazy chunk.
  • react-markdown gets a 1.1 KB group of its own, so it does not land on the budgeted ui-components line. fields/MarkdownContent.tsx imports it eagerly (objectui#5325).

• Deviation, accepted: the dispatch named a separate lazy vendor-markdown. The lazy remainder instead rides plugin-markdown's own chunk, which saves the docs page one request. The direction asked that the unreached modules leave the first load, and they do.

• Ledger lowered by exactly what left: the eager closure falls from 3,433,602 to 3,375,344 bytes gzipped (−58,258). MAX_EAGER_CLOSURE_GZIP_BYTES falls from 3,444,622 to 3,386,364, the same delta, so headroom stays at main's 10.8 KB. BASELINE is re-pinned to the measured commit.

• Not done here: 104,286 bytes stay in the eager vendor-markdown, because streamdown, imported statically by the eager plugin-chatbot, reaches them on every page. Moving them means lazy-loading a reached module: objectui#11798's question, with the maintainer under the objectui#6795 ruling.

• No chunk cycle (H2): a Tarjan pass over the chunk import graph finds one SCC on both builds, the pre-existing data-adapter / framework / i18n-runtime one recorded in the config, and none involving a markdown chunk.

• Real-browser control on both builds, in Chromium:

  • the first load fetches exactly the eager closure (333 → 334 chunks), and plugin-markdown is not in it;
  • the docs page renders highlighting, slugs, autolinks, a GFM table and an alert;
  • a shared chatbot reply renders through streamdown;
  • zero page errors, zero console errors, zero failed dynamic imports.

  The markdown-field path was not reached (no record fixtures). Its react-markdown is the same chunk the docs page exercised.

• Clause-②: no holds: only which chunk a module ships in changes. No entry, route, registry key or translation key changes.

• Changeset: @object-ui/console: patch. Every sentence was checked against the diff.

• Evidence:

  • check:eager-closure passes at 3296.2 KB against 3307.0 KB.
  • Reverse check on the gate: the re-pinned gate on the base build exits 1, 46.1 KB over. Reverse check on the unit: removing the tag turns exactly the $initial control red. The restore was proven.
  • 24 files and 1,114 tests pass: every test that reads the console config or the gate module.

• CI on 8027e1e: no failure so far, 13 checks still running. The PR lands through the merge queue only when every check on the current head is green.

Out-of-scope finding: none filed. Every console build prints rolldown's "advancedChunks option is deprecated" notice. That is deprecation drift, carrier: none.

---

Generated by Claude Code