Beautiful, AI-native markdown IDE and LLM wiki
Bundle-size limits — rationale and history
The size-limit entries in package.json are hand-maintained: size-limit ships
no auto-update or ratchet, so each limit is measured and then edited by hand. This
file records why each limit moved, so that reasoning survives without living on a
single line of package.json that every branch has to append to.
The limits themselves stay in package.json — they are read by the
open-knowledge / size CI job. Only the narrative moved here.
Open follow-ups
TODO: check why after installing @pierre/trees has exceeded by 55 kB.
- PropertyPanel pulled date-fns + react-day-picker into the eager bundle (~24 kB) — follow-up should lazy-load DateWidget so the calendar/date-fns chunk only loads on first date-property edit.
- @dnd-kit adds ~15 kB to the eager bundle (PropertyPanel imports it at top level) — follow-up should lazy-load PropertyPanel itself so dnd-kit + the widget zoo only load when a doc has frontmatter.
- terminal multi-session added ~2.65 kB to the eager bundle (TerminalTabStrip + the Radix Tabs primitive) — follow-up should lazy-load TerminalDock, which is desktop-only (gated on terminalBridge) and ships as dead weight in the web bundle.
- the bottom Ask AI composer (BottomComposer + AgentSplitButton) sits in the eager bundle though it is desktop-only (gated on isDesktop) — follow-up should lazy-load BottomComposer so it stops shipping as dead weight in the web bundle (would reclaim ~4 kB).
Limit-change history
Limit raised 410->450 kB after merging main + the skill/template editor feature; the skill-specific UI (ManagedArtifactProperties, SkillEditorActions) is already lazy-loaded, so the residual is the pre-existing PropertyPanel tech-debt above — lazy-loading PropertyPanel should bring this back under 410.
Main CSS limit raised 49->52 kB after merging main + the skills-as-content UI — Tailwind utility accretion across many new surfaces, no single bloat source and no lazy-load lever for global CSS.
Limit raised 450->454 kB after merging main + the worktree selector feature (base-branch picker with local/remote search, flyout keyboard nav, project-navigator icon tiles); ~2 kB of this bump is concurrent main growth (incl. terminal-chat-controls #2350) merged in during finalization, and the PropertyPanel/TerminalDock/BottomComposer lazy-load levers above remain the reclamation path to bring this back down.
Combined-chunks limit raised 3.05->3.06 MB after adding the standalone Mermaid doc editor (MermaidDocEditor: diagram + editable Y.Text source with mermaid highlighting) — it is lazy-loaded (main bundle unaffected, still under 460 kB) so the growth is a new lazy chunk, not eager weight; the mermaid renderer + codemirror-lang-mermaid grammar were already in the chunk graph (prop-editor + edit-modal), so the delta is the doc-editor glue only.
Raised again 3.06->3.07 MB when rebasing onto main during finalization — ~2 kB is concurrent main growth (main app bundle grew 456.8->458.1 kB from unrelated merges), not the mermaid editor.
Main app bundle limit raised 460→462 kB after the mermaid-doc Ask-AI-composer-clearing fix (78ff0e5) pushed eager weight to 460.33 kB (333 B over) — the BottomComposer lazy-load lever above remains the reclamation path.
Combined-chunks limit raised 3.07->3.08 MB after the terminal clickable-links feature (terminal-links + terminal-link-provider land in the already-lazy TerminalPanel chunk, not the eager bundle — main app stayed at 460.84 kB) plus concurrent main growth on rebase pushed the combined total 315 B over 3.07 MB; the TerminalDock lazy-load lever above remains the reclamation path.
Main app bundle limit raised 462→466 kB after merging main + the tree-view Show visibility feature: eager weight is the Show menu group across the tree-options popover / context menu, the shared visibility predicates, and the NotInSidebarIndicator (~1-2 kB), with the rest concurrent main growth absorbed across three catch-up merges (bundle measured 464.63 kB, 2.63 kB over); the PropertyPanel/TerminalDock/BottomComposer lazy-load levers above remain the reclamation path.
Combined-chunks limit raised 3.08->3.25 MB (measured 3.22 MB after merging main) for the ACP agent-threads feature (in-app threads + unified sessions dock): the thread-client/thread-event-model chain, ThreadView, and the dock UI are lazy chunks (TerminalSessionsHost, AgentBuildShowcase, AgentThreadClientBinder, and the desktop TerminalWindowApp are all React.lazy mounts, so the main bundle stays under its 466 kB limit at 458 kB), but the combined glob counts lazy chunks too, so the feature's total JS footprint lands here.
Main CSS limit raised 52->53 kB — merging the ACP dock/thread surfaces onto main's CSS accretion pushed global utilities to 52.86 kB (863 B over), no single bloat source and no lazy-load lever for global CSS.
Combined-chunks 3.25->3.35 MB and CSS 54->56 kB after merging main (markdownlint content rules, Themes, external-link-preview, in-app Report a bug, GitHub Copilot CLI) into the ACP-threads branch — the merged build measures 3.31 MB / 55 kB combining both feature sets (main app 469 kB, under its 476 kB limit); no single reclaimable source beyond the lazy-load levers above.
Combined-chunks 3.35->3.5 MB and CSS 56->57 kB after merging main's bun->pnpm (Node 24) toolchain migration into the ACP-threads branch — a fresh pnpm resolution floats several bundled deps to newer patches within their existing caret ranges (dep declarations unchanged, no duplicate versions), diffuse drift inherent to the package-manager swap with no single reclaimable source; measured 3.45 MB / 56.38 kB, main app stays 469 kB under its 482 kB limit.
Merged the in-app feedback form onto these post-toolchain limits: main-app 482->485 kB (feedback's only eager weight is the HelpPopover Provide-feedback entry + dialog shell; FeedbackForm and its zod/react-hook-form/attachment deps are lazy behind FeedbackFormDialog), CSS 57->59 kB (the form's globals.css utilities — .scroll-fade-x mask, .scrollbar-none, shimmer @utility/@keyframes — plus Tailwind accretion from FeedbackForm + the attachment card), combined 3.5->3.55 MB (the merged tree stacks main's ACP/toolchain features and the lazy FeedbackForm chunk — measured 3,465,453 B locally, ~3.51 MB projected in CI at the ~1.4% local->CI ratio, over the intermediate 3.5); CI size-limit verifies the true merged numbers.
Windows/Linux desktop port stacked on this: the win/linux-only AppMenubar (and its radix Menubar primitive) is lazy-loaded behind the app-menubar-gate predicate so web + macOS never download its chunk; the residual eager weight is the gate + lazy() glue plus the pty-capability gating in EditorPane/SettingsDialogShell, absorbed under the current limits with no reclaimable source beyond the pre-existing lazy-load levers above.
Combined-chunks limit raised 3.6->3.61 MB for the div/center align-wrapper rendering (div-align promoter + HtmlAlignBlock descriptor/renderer; measured 170 B over the combined budget, main bundle unaffected at 485 kB) — the PropertyPanel/TerminalDock/BottomComposer lazy-load levers above remain the reclamation path.
Merged the skills-marketplace branch (in-place skill model, location menus, folder verbs, MCP location parity) onto these post-toolchain limits; both feature sets stack and CI size-limit verifies the true merged numbers. Combined-chunks 3.61->3.7 MB after two catch-up merges of main into this branch (post-stable v0.41.3 and v0.41.4) stacked on the divergent-same-name-skill rows, the per-location convert verb, and the regrouped folders panel — CI measured 3.65 MB (53.55 kB over 3.6); the main app bundle is unaffected at 377 kB against its 485 kB limit, so the growth is lazy-chunk and concurrent-main weight rather than eager, and the PropertyPanel/TerminalDock/BottomComposer lazy-load levers above remain the reclamation path.
Combined-chunks 3.7->3.72 MB after merging main (in-doc comments: CommentedDocProvider / PropertyCommentButton / the comment queue + margin rail, plus the derived-index onDocumentSettled hook) into the skills-marketplace branch. CI measured 4.14 kB over 3.7 MB; both feature sets stack and neither is the single source. The main app bundle is unaffected at 395 kB against its 485 kB limit and CSS at 8.4 kB against 59 kB, so the growth is lazy-chunk weight rather than eager — the PropertyPanel/TerminalDock/BottomComposer lazy-load levers above remain the reclamation path.
Combined-chunks limit raised 3.78->3.95 MB (on the ACP bug-bash quick-wins branch; superseded by the parallel 4.5 MB raise below before reaching main) for ACP syntax highlighting: the in-house streamdown code-highlighter plugin — shiki core + the JavaScript regex engine (no oniguruma wasm) + 15 curated grammars, with the TS family collapsed onto tsx/javascript since the four near-copy grammars cost ~140 kB gz for no visible difference — lands entirely in a lazy chunk loaded on the first fenced code block. Measured 3.91 MB combined; the main app bundle is unaffected at 408 kB against its 487 kB limit. The off-the-shelf @streamdown/code plugin (full-bundle shiki, every grammar) measured 5.57 MB and was replaced by the curated plugin; trimming the grammar list further is the reclamation lever.
Combined-chunks 3.78->4.4 MB when the interface-language preference started activating catalogs. The ten non-English locales become real on-demand chunks the moment ConfigProvider asks for one; until then nothing imported the loader map and rolldown dropped it, so the chunks did not exist to be measured. They add 580,340 B gzipped, 58,034 B each — every catalog is full-size regardless of how much of it is translated, because an untranslated entry compiles to its English source string, so the skeleton catalogs already weigh what finished ones will. Measured 4.31 MB. The main app bundle is untouched at 395.62 kB against its 485 kB limit and CSS at 8.36 kB against 59 kB, which is the point of splitting them: a reader downloads one catalog, not eleven. Importing all eleven eagerly would have put ~638 kB gzipped in front of first paint, more than the entire main-bundle budget on its own, so there is no reclamation lever here — the weight is the feature.
Combined-chunks 4.4->4.5 MB once every catalog held real translations rather than English skeletons. Measured 4.41 MB, 6.02 kB over. The entry above predicted this would cost nothing: an untranslated entry compiles to its English source string, so a skeleton already weighs a full catalog, and translating it swaps bytes rather than adding them. That holds for Latin and CJK and does not hold for the Indic and Arabic scripts. Measured per-locale gzipped chunks now span 60,026 B to 78,840 B, where before every one sat at 58,034 B: Devanagari and Bengali are the two largest at ~78 kB, Arabic and Urdu next at ~71 kB, and the Latin and Han catalogs stay near 60-65 kB. Those scripts are 2-3 UTF-8 bytes per character against English's 1, and gzip recovers less of that on an alphabet as large and combining-mark-heavy as Devanagari than it does on Han. The main app bundle is untouched at 395.23 kB against its 485 kB limit and CSS at 8.36 kB against 59 kB, so this is entirely on-demand catalog weight: a reader still downloads one catalog, and the one they download is now the one they can read. No reclamation lever — same as above, the weight is the feature.
Combined-chunks 4.5->4.7 MB merging the ACP-highlighting branch onto the language catalogs: the two lazy features stack — eleven on-demand locale catalogs plus the curated shiki grammar chunk — and measured 4.6 MB locally, ~4.66 MB projected in CI at the ~1.4% local->CI ratio, 104 kB over the 4.5 ceiling that was set before the highlighter existed. The main app bundle and CSS stay under their limits; both features load nothing until used (a catalog on locale switch, the grammar chunk on the first fenced code block). Reclamation levers: trim the grammar list, or the catalog-weight notes above.
Combined-chunks 4.7->4.71 MB for the frontmatter-schema folder picker: the searchable multi-select UI, live glob-match diagnostics, and their translations across all supported locale chunks measured 4,701,170 B in CI, 1,170 B over the prior ceiling. The eager app bundle remains 415.93 kB against 487 kB and CSS remains 8.36 kB against 59 kB; this is a measured 10 kB combined-budget increment for the shipped feature rather than an eager-loading regression.
Combined-chunks 4.7->4.72 MB for uniform local-target classification and image failure states across the editor, Problems, Links, and source diagnostics. CI measured 4.71 MB, 8.53 kB over the prior ceiling, while the main app bundle remained at 418.08 kB against its 487 kB limit and CSS at 8.36 kB against 59 kB. The increase therefore records the feature's total chunk footprint rather than first-load growth; the pre-existing lazy-load levers above remain the reclamation path.
Combined-chunks kept at 4.75 MB when merging main. Both sides had raised the same budget off a 4.7 MB base that had no headroom left: this branch to 4.75 for the gitignored-skill dialog (measured 364 B over, 0.008%), main to 4.72 for the entry above. The merged tree carries both features, so the resolution takes the higher ceiling rather than either side's number. CI measured 4.7 MB against 4.75 on the pre-merge branch head and the post-merge tree measures 4.69 MB locally, where local reads a little under CI. The lazy-load levers above remain the reclamation path.
Combined-chunks 4.75->4.78 MB for the Open Knowledge Format plugin controls, generated-index settings, starter-pack install card, and their translations across all supported locale chunks. CI measured 4.76 MB, 12.84 kB over the prior ceiling, while the main app bundle remained at 417.92 kB against its 487 kB limit and CSS at 8.36 kB against 59 kB. The increase records the feature's total chunk footprint rather than first-load growth; the pre-existing lazy-load levers above remain the reclamation path.
Combined-chunks 4.85->4.95 MB for Korean (ko), the twelfth supported locale. Its catalog is emitted as a separate on-demand chunk, so the increase is combined-chunk inventory rather than eager weight: the branch measured 4.86 MB locally, 8.85 kB over the prior ceiling, while the main app remained at 420.63 kB against its 487 kB limit and CSS at 8.36 kB against 59 kB. The new ceiling leaves about 91 kB (1.9%) of headroom for the local-to-CI variance recorded above.
Combined-chunks 4.95->6.55 MB for the .excalidraw canvas doc editor (@excalidraw/excalidraw@^0.18.1, matched to Excalidraw's own file format via serializeAsJSON / restore). Weight is entirely on-demand: ExcalidrawDocEditor mounts through EditorActivityPool behind the existing React.lazy gate, and Excalidraw itself defers its scene renderer, mermaid-to-excalidraw importer, and rough.js / perfect-freehand stack into its own async chunks — the eager main bundle is unaffected (measured 421.2 kB against its 487 kB limit). The gzipped total across the added chunk-EIO257PC-*.js + dist-*.js + ExcalidrawDocEditor-*.js chunks measures 6.41 MB locally, ~1.46 MB above the prior 4.95 MB ceiling; 6.55 MB carries the addition with ~140 kB headroom for the local-to-CI variance recorded above. No smaller Excalidraw slice exists on npm (the package ships one entry with the scene renderer, roughjs, perfect-freehand, @excalidraw/laser-pointer, and @excalidraw/mermaid-to-excalidraw all wired together), and the workspace deliberately vendors the fonts to packages/app/public/excalidraw-assets/ rather than accepting the upstream esm.sh CDN fallback (see scripts/copy-excalidraw-assets.mjs) — the vendored assets are static files under public/, not JS, so they do not count against this budget.
Combined-chunks 6.55->6.56 MB for reload-safe terminal consent across config-epoch rebuilds. The gate retains the last fully synced project-local value so a live adopted PTY is not unmounted during the transient unsynced window, while an explicit opt-out remains closed. Exact CI measured 6,550,051 B, 51 B over the former ceiling; the eager main bundle remained at 438.61 kB against 487 kB because the terminal session surface is lazy. The 10 kB ratchet records the measured state-machine cost without changing the eager-load budget.
The <Excalidraw src> document embed adds a second reachability path into the same chunk: it lazy-imports @excalidraw/excalidraw on first embed mount, so the download now also fires when a markdown document containing the block renders — a passive document view, not just the explicit act of opening a .excalidraw doc. The bytes dedupe into the same chunk (byte-neutral for the budget); only when they arrive changed.
Combined-chunks 6.56->6.6 MB after the embed's review remediation landed CI at 6,550,489 B — 489 B over the pre-terminal-consent ceiling. The delta is small and diffuse (the embed's retry/provenance handling, the shared live-doc pool gates, and new error strings compiled into all 13 locale catalogs, which ship as JS chunks); the prior local measurement's assumed headroom did not survive the CI toolchain's chunking. 6.6 MB carries both measured additions (the embed's and the terminal-consent ratchet merged in from main) with ~49 kB of local-to-CI variance headroom rather than a ceiling calibrated to a single local run.
Combined-chunks 6.6->8.4 MB for @pierre/[email protected], the merge-conflict and diff renderer replacing @codemirror/merge. The weight is one dependency's transitive shiki grammar set, not the feature: dist/highlighter/languages/resolveLanguage.js statically imports bundledLanguages from shiki, so Vite emits a chunk per grammar and the build gains roughly 310 of them — apl, blade, bsl, cairo, and so on. Measured against a main build: all JS chunks 6.22 -> 7.94 MB gz, chunk count 464 -> 775. None of it is delivered. Every grammar chunk is dynamic-import-only, the entry HTML preloads the same 179 files it did before, and a session downloads 2.04 -> 2.17 MB gz.
The eager budgets did not move, and that is the load-bearing fact. Main app bundle stays at 487 kB and CSS at 59 kB. The first revision of this work did put Pierre in the eager chunk — every surface was statically imported from always-mounted chrome, measuring 459.28 kB — which was then code-split behind the boundaries that already gate those surfaces (DiffViewBoundary on lifecycle.status, the two Skill dialogs on user action). Measured after: 437.42 kB against main's own 438.38 kB, so the swap costs the eager bundle nothing.
Transitive additions worth attributing: @pierre/[email protected], @pierre/[email protected], @shikijs/transformers, diff@9, lru_map arrive while @shikijs/themes leaves dependencies, so the gross Pierre cost exceeds the 1.85 MB delta. Reclamation lever: alias the bare shiki specifier to a shim exporting bundledLanguages = {} and register only the grammars a diff needs through Pierre's RegisteredCustomLanguages. The app's own highlighter imports shiki/core plus individual @shikijs/langs/* entries and never the full-bundle entry, so such a shim would not touch it. Deliberately not done here: it caps which languages highlight inside a diff, and the bytes it reclaims are ones no session downloads.
Combined-chunks 8.4->8.8 MB for the @excalidraw/excalidraw bump to 0.18.0-abeeaeb, the release-branch snapshot that carries the stroke Pressure setting. Measured against the merged tree: 8,588,202 B locally, ~8.71 MB projected in CI at the ~1.4% local->CI ratio, roughly 190 kB over the 8.4 ceiling. An intermediate 8.65 was set first and would have landed red — it clears the local number but not the projection, which is the reason this file records the ratio at all. The snapshot is about 99 upstream commits ahead of 0.18.1, and the delta is not attributed here. It is not new third-party weight: @excalidraw/[email protected], which the constant-width renderer draws through, was already a declared dependency of 0.18.1, so the Pressure work adds first-party renderer code rather than a library. Ninety-nine commits is a wide enough window that pinning the bytes on any one of them would need a per-commit build, which was not run.
The eager budgets did not move. The main app bundle measures 437.95 kB against its 487 kB limit and CSS 8.36 kB against 59 kB, because both surfaces that load a board — the standalone .excalidraw editor and the in-document snapshot embed — still reach the package through a dynamic import. Nobody downloads any of this unless they open a board. Reclamation lever: none worth taking here; the weight is upstream's own renderer and the chunk is already gated behind the two board surfaces. Headroom at the new ceiling is 90 kB against the CI projection, in line with the margins above.
Combined-chunks 6.55->6.56 MB for the merged agents settings page. CI measured 6,550,218 B, 218 B over the prior ceiling, while the main app bundle stayed at 438.42 kB against its 487 kB limit and CSS at 8.36 kB against 59 kB — so the growth is chunk inventory rather than eager weight. Base-vs-branch was measured rather than assumed: 2.4 kB gzipped) and its new strings across the twelve locale catalogs (~4.1 kB, consistent with the per-catalog costs recorded above), with the remainder chunk reshuffle. Note the 218 B is a CI measurement, not a local one, so the ~9 kB the new ceiling adds is real headroom rather than variance cover. Raising was taken over trimming copy the page needs; the lazy-load levers above remain the reclamation path.feat/agent-integration-requirements-matrix builds under the prior ceiling and this branch does not, so the increase is this branch's and not inherited headroom loss. It splits into the page itself (
Combined-chunks 6.56->6.57 MB while hiding connection options an agent has no surface for. The branch's own contribution is small and the ceiling was already breached without it: measured base-vs-branch on one machine at merge base 3372c64, feat/agent-integration-requirements-matrix alone builds to 6,564,728 B — 4.73 kB over the 6.56 MB ceiling it ships with — while this branch builds under 6.57 MB. So the raise carries growth that arrived on the base (the merged agents-settings work plus its new strings across the twelve locale catalogs), not primarily this branch's dialog change. CI measured this branch at 6,560,777 B before review cleanups and 6,560,730 B after, against the prior 6,560,000 B ceiling. The main app bundle is byte-identical between base and branch at 438.68 kB against its 487 kB limit, and CSS at 8.36 kB against 59 kB, so the growth is chunk inventory rather than eager weight. Recording the base measurement explicitly because the earlier note on this branch asserted the overage was the branch's own — that was inferred from it being the only app-code change, not measured, and the measurement contradicts it. The lazy-load levers above remain the reclamation path, and the base being over its own ceiling is worth an owner independent of this branch.
Combined-chunks 8.8->8.82 MB for the searchable chat-history side panel, its responsive presentation state, and the new strings compiled into every locale catalog. CI measured 8,804,992 B against the prior 8,800,000 B ceiling; the eager main bundle remained at 437.22 kB against 487 kB and CSS at 8.36 kB against 59 kB. The 20 kB ratchet records the feature's total chunk inventory with about 15 kB of headroom while preserving the existing eager-load budgets.
Combined-chunks 8.84->8.85 MB for GitHub Enterprise Server host handling in Settings → Account: the hosts card (list, add field, input parsing) and the explanation shown when a project's remote is not a recognized GitHub host, plus their 17 new strings compiled into every locale catalog. CI measured 8,841,175 B, 1,175 B over the prior ceiling; a local build measured 8,840,090 B. The main app bundle stayed at 438.19 kB against its 487 kB limit and CSS at 8.36 kB against 59 kB, so the growth is on-demand chunk weight: the settings dialog body is lazy-loaded and each reader downloads one catalog. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 8.84->8.86 MB for repository-aware worktree discovery and presentation. A forced clean local build measured 8,840,825 B against the prior 8,840,000 B ceiling, with 26,938 B of growth over the same-machine main build at 8,813,887 B. The eager main-app and CSS budgets remain unchanged at 487 kB and 59 kB because this ratchet records the feature's aggregate chunk inventory rather than a new first-load allowance. The new ceiling leaves 19,175 B of headroom over that local measurement; the corresponding CI size job passed.
Combined-chunks 8.88->8.89 MB for GitHub reference links and hover cards in agent chat: owner/repo#N and configured ticket-prefix linking, the hover-card primitive, GitHubReferenceCard with its status line, and their strings compiled into every locale catalog. CI measured the pull request at 8,881,800 B, 1,800 B over the prior ceiling. Measured on one machine after the review fixes, main at 869755a builds to 8,868,939 B and the same main with this branch merged builds to 8,881,731 B, so the branch adds 12,792 B. The main app bundle moved 36 B, to 441.88 kB against its 487 kB limit, and CSS stayed at 8.36 kB against 59 kB, because the chat that renders these links and the settings dialog body are already lazy chunks; the growth is on-demand chunk weight. The 10 kB ratchet leaves 8,269 B of headroom over the merged local build, and the lazy-load levers above remain the reclamation path.
Combined-chunks 8.89->8.90 MB after merging the comments-panel layout and selection polish with the newer main budget. The merged app measured 8,890,802 B locally, 802 B over the 8,890,000 B ceiling; the earlier PR CI build on the old base measured 8,890,823 B. The main app bundle measured 442.71 kB against 487 kB and CSS stayed at 8.36 kB against 59 kB. The 10 kB ratchet records the combined comments UI and newer-main growth without relaxing either eager-load budget. Built with the same toolchain on this machine, current main measures 8,889,952 B and the merged comments branch measures 8,890,806 B, an 854 B contribution from this branch. The 8,900,000 B ceiling leaves 9,194 B of headroom.
Combined-chunks 8.89->8.9 MB for PowerShell path quoting: psQuoteArg now doubles all five characters PowerShell reads as a single quote (U+0027 and U+2018-U+201B), not only the ASCII apostrophe, so a dropped file whose name contains a curly quote gives a working command. CI measured the pull request at 8,890,002 B, 2 B over the prior 8,890,000 B ceiling. Clean local builds of the same tree measured 8,889,973 B with the previous psQuoteArg and 8,890,002 B with the change, so it adds 29 B. The main app bundle measured 441.88 kB before the change and 441.87 kB with it, against its 487 kB limit, and CSS stayed at 8.36 kB against 59 kB, so neither eager budget moves. The 10 kB ratchet leaves 9,998 B of headroom over the local build.
Combined-chunks 8.9->9.15 MB for Polish (pl), the thirteenth supported locale. Base-vs-branch was measured rather than assumed: with pl removed from the locale tuples and rebuilt on this branch, combined chunks measured exactly 8.9 MB — on its ceiling with no headroom left — so the whole overage is this branch's. With pl enumerated the same build measures 8.99 MB, 94.09 kB over the 8,900,000 B ceiling, and the Polish catalog's own on-demand chunk measures 96,360 B gzipped, which accounts for the increase. The main app bundle is byte-comparable between the two builds at 443.24 vs 443.25 kB against its 487 kB limit, and CSS at 8.36 kB against 59 kB, so this is chunk inventory rather than eager weight: a reader who never selects polski never fetches it. The new ceiling leaves about 156 kB (1.7%) of headroom, in line with the 1.4% local-to-CI ratio and the ~1.9% recorded on the ko raise above.
Combined-chunks 8.9->8.91 MB for the header pane buttons: the Editor, Agents and Terminal toggles, the Nothing open screen, the reveal-on-deliberate-open plumbing, the Show/Hide Editor command, and their strings compiled into every locale catalog. CI measured the pull request at 8,902,839 B, 2,839 B over the prior 8,900,000 B ceiling, which the base feat/layout-redesign already sat on. The main app bundle grew from 448.28 kB on the base to 453.01 kB against its 487 kB limit, because the header buttons, the pane store and the reveal helpers load eagerly with the shell; CSS stayed at 8.36 kB against 59 kB. The 10 kB ratchet leaves 7,161 B of headroom over the CI measurement. Removed on the same branch: the right-edge agents tab and the in-pane collapse chevrons, which is why the combined growth is smaller than the eager growth. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 8.91->8.92 MB for the corner card stack that floats update and feedback cards in the window's bottom-left corner. CI measured the pull request at 8,911,181 B, 1,181 B over the 8,910,000 B ceiling; the same job on the base feat/layout-redesign head 0c8e62f measured 8,910,714 B, already 714 B over, so this branch adds 467 B. The main app bundle measured 453.61 kB on the base and 453.8 kB on the branch against its 487 kB limit, and CSS stayed at 8.36 kB against 59 kB. The 10 kB ratchet leaves 8,819 B of headroom over the CI measurement. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 8.92->8.93 MB for the responsive single-pane workspace: one selected component in narrow windows, the sidebar covers over the editor card, the shared sidebar fit budget, the narrow editor group selector, the single-file shared shell, and their strings compiled into every locale catalog. CI measured the pull request 8.48 kB over the 8,920,000 B ceiling, about 8,928,480 B. The main app bundle measured 467.58 kB against its 487 kB limit, up from 453.8 kB on the base, because the narrow layout state and the cover controls load eagerly with the shell; CSS stayed at 8.36 kB against 59 kB. The 10 kB ratchet leaves about 1.5 kB of headroom over the CI measurement. No reclamation lever beyond the lazy-load levers above.
Combined-chunks stays at 9.15 MB after merging feat/layout-redesign into main. The three raises above (8.9->8.93 MB) were set on that integration branch while main raised the same ceiling to 9.15 MB for Polish, so the enforced limit is main's and the 8.91, 8.92 and 8.93 MB ceilings never applied there; their headroom figures describe the branch, not the merged tree. A clean local build of the merged tree measured 9,144,866 B combined, 5,134 B under the 9,150,000 B ceiling, with the main app bundle at 477.79 kB against its 487 kB limit and CSS at 8.36 kB against 59 kB. CI's size job passed on the first merged head. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 9.15->9.16 MB for @visimer/core and @visimer/dom 1.x, which split the Mermaid edge picker into separate arrow, stroke, color, animation and curve pickers. CI measured the bump at 9,150,517 B, 517 B over the 9,150,000 B ceiling. Direct local vite build runs of packages/app measured 9,146,208 B with 0.1.0 and 9,149,900 B with 1.1.2, so the bump adds 3,692 B. CI measures about 600 B more than a local build, so main sat about 3.2 kB under the old ceiling before this change. The main app bundle measured 478.46 kB against its 487 kB limit, and CSS stayed at 8.36 kB against 59 kB. The 10 kB ratchet leaves about 9.5 kB of headroom over the CI measurement. None of that headroom is reserved: the HTML preview frame page, web shell CSP and desktop file protocol changes that land after this one pass under 9.16 MB with no further raise. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 9.16->9.17 MB for the quiet notice after a both-kept merge or a version moved to history (PRD-9735). Measured 9,157,371 B on main and 9,165,865 B with the notice, 8,494 B more: about 6.6 kB is its 14 new strings in each of the 13 locale catalogs, about 1.7 kB the notice strip, its data hook and the timeline row labels in the main bundle (481.29 kB against its 487 kB limit), and 0.15 kB the notice schema in the schemas/api chunk, which the app already imports through that narrow entry. No reclamation lever: the strings are the feature, and each reader downloads one catalog.
Combined-chunks 9.17->9.18 MB for keeping the edited side of engine-less files, settings and skills (PRD-9771, #7203): the history and backup-folder notices in DocChangeNotice, the settings notices, the held-save message, and their new strings compiled into each of the 13 locale catalogs. Measured on one machine, main alone builds to 9,166,292 B and the branch to 9,175,290 B, so the branch adds 8,998 B and lands 5,290 B over the prior 9,170,000 B ceiling. The 10 kB ratchet to 9,180,000 B leaves 4,710 B of headroom over the branch build. No reclamation lever: the strings are the feature, and each reader downloads one catalog.
Combined-chunks 9.18->9.19 MB for moving docked chat history to either side of the conversation (#7294): the Agents strip's row menu with its Move chat history item, the palette's side-aware move commands, and their strings compiled into each of the 13 locale catalogs. Measured on one machine, main at 450b65e alone builds to 9,178,809 B, 1,191 B under the 9,180,000 B ceiling, and the branch merged with it to 9,180,163 B, so the branch adds 1,354 B and lands 163 B over; the merge queue measured the same 9,180,163 B. The main app bundle went from 482.47 kB on main to 482.26 kB on the branch against its 487 kB limit, and CSS stayed at 8.36 kB against 59 kB, so the growth sits outside the main app bundle. The 10 kB ratchet to 9,190,000 B leaves 9,837 B of headroom over the branch build. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 9.19->9.20 MB for the attachment folder field with {{filename}}: one path input with its variable highlight, the live example and its errors, the zod input schema and per-document folder helpers that the field shares with the server, the translated upload refusal copy, and their strings compiled into every locale catalog. The chat-history entry above measured the tree before this field at 9,180,163 B. Merged with main at 33ac0b4, the branch measures 9,191,157 B locally, 1,157 B over the 9.19 MB ceiling. The 10 kB ratchet leaves 8,843 B of headroom locally, or about 8.2 kB with the ~600 B that CI adds over a local build. The main app bundle measured 482.29 kB against its 487 kB limit and CSS 8.36 kB against 59 kB, because the settings dialog body is a lazy chunk. No reclamation lever beyond the lazy-load levers above.
Combined-chunks 9.20->9.21 MB for the Zod upgrade from 4.4.3 to 4.6.7 (#7691). The app takes only the larger Zod core, because the zod/compile side effect loads in the CLI and the desktop server process, never in the browser. Measured on one machine, main at 8fc6a56 builds to 9,195,153 B and the branch merged with it to 9,202,392 B. The upgrade adds 7,239 B and lands 2,392 B over the 9.20 MB ceiling, and CI measured the same 9,202,392 B. The shared schemas chunk that holds Zod grows 7,304 B, and the other chunks move by a few bytes each. The main app bundle moved 2 B to 482.86 kB against its 487 kB limit, and CSS stayed at 8.36 kB against 59 kB. The 10 kB ratchet leaves 7,608 B of headroom. No reclamation lever beyond the lazy-load levers above.