[maartenbreddels]

The Jupyter output renderers leave the core bundle, so pages that leave out output-widget, like a minimal page without an Output widget, skip them. In production builds, the core shrinks by about 33 KB gzip on Vue 3 and 29 KB on Vue 2. The Vue 3 number includes two Vue build flags. This PR is now the bottom of a stack on #1235. jQuery and Lumino stay in the core here, and move out in the lumino PR (#1240) and the jquery PR (#1241).

Problem

Every Solara page downloads and runs one core bundle before it shows the first widget.
After #1235, this core is still 244 KB gzip on Vue 3 and 359 KB on Vue 2 (ipywidgets 8, production build).
About 29 KB of it is the Jupyter output renderers: the code that shows HTML, Markdown, LaTeX, images and JavaScript outputs.
Only the ipywidgets Output widget (for example solara.display of an IPython object) uses them.
Most pages have no Output widget, but every page pays to download and run this code.
Slow phones and laptops pay the most.

The Vue 3 core also keeps code for two Vue build flags that Solara never changes, because the build does not tell Vue their values.

Change

• The --frontend setting of feat: load pages faster by splitting the frontend into optional features #1235 picks the frontend features that a page preloads.
  The preset full (the default) preloads all of them.
  The preset minimal leaves most of them out, and the browser loads those on first use, with a warning.
• The output renderers move into the output-widget chunk.
  These are @jupyterlab/rendermime with marked, the HTML sanitizer of @jupyterlab/apputils, @jupyterlab/translation, and the JavaScript renderer.
• The core keeps a small RenderMimeRegistry stand-in.
  It remembers its options and the renderers added to it.
  When the output-widget chunk runs, the stand-in makes a real registry with the same options and renderers, and passes every call on to it.
• full preloads output-widget.
  The chunk then runs before the page makes its registry, so a full page gets a real registry, as in feat: load pages faster by splitting the frontend into optional features #1235.
• In minimal, the first Output widget loads output-widget, as in feat: load pages faster by splitting the frontend into optional features #1235.
  Before that, the stand-in accepts only addFactory, clone and its option fields.
  Any other use, for example createRenderer, getFactory or a read of mimeTypes, throws an error and starts the load.
  The docs now say this for a rendering call such as createRenderer.
• Like the real registry, the stand-in drops a renderer that a later renderer fully replaces.
  A soft remount happens when a page reconnects to a new kernel: the page makes a new widget manager and renders the app again, without a reload.
  Each soft remount adds a renderer that holds the new widget manager.
  Without this fix, a page that never loads the Output widget keeps every old widget manager in memory.
• The Vue 3 build defines __VUE_OPTIONS_API__ (on) and __VUE_PROD_HYDRATION_MISMATCH_DETAILS__ (off), their runtime defaults.
  The minifier then removes the runtime checks for both flags, and the code that the hydration mismatch details flag turns off.
  The Options API stays on.
  __VUE_PROD_DEVTOOLS__ stays a runtime global, so a page can still turn on the Vue devtools in production.
• This version of the PR is smaller than the earlier version (commit b17a9bcb):
  • jQuery and Lumino move to the lumino PR (perf: let minimal pages skip most of Lumino until a widget needs it #1240) and the jquery PR (perf: let minimal pages without jQuery widgets skip loading jQuery (opt-in) #1241), so the classic-widgets feature is gone.
  • The semver port and its license commit are dropped.
  • The stand-ins for @jupyterlab/coreutils, the kernelspec API and minimist are dropped, and so is their build check, StandInGuardPlugin.
  • The create_view hook, the prototype swap of $el objects and the list of guessed modules are dropped.
  • Q45 and Q47 (see Align results) settled these changes.
  • In an earlier measurement, the earlier version had a Vue 3 core of about 129 KB gzip. This PR has 212 KB, and the top of the stack has 142 KB.

Validation

I built all 8 bundles once: Vue 2 and 3, ipywidgets 7 and 8, production and development.
I compared them with the build of #1235 at 60586ce8.
Sizes are bytes after gzip -9 -n.
"Full JS" is the core plus every chunk that the full preset preloads.

build	core gzip, #1235	core gzip, this PR	output-widget gzip	full JS gzip
Vue 3, ipywidgets 8, prod	244,341	211,539	9,442 to 39,431	610,763 to 607,910
Vue 3, ipywidgets 7, prod	239,140	206,203	9,445 to 39,626	596,589 to 594,021
Vue 2, ipywidgets 8, prod	358,887	329,787	9,440 to 39,429	529,563 to 530,452
Vue 2, ipywidgets 7, prod	353,632	324,386	9,444 to 39,625	530,522 to 531,686
Vue 3, ipywidgets 8, dev	643,542	586,152	27,336 to 86,674	1,359,633 to 1,361,553
Vue 2, ipywidgets 8, dev	835,829	778,631	27,335 to 86,672	1,189,575 to 1,191,714

• In production builds, the renderers save about 29 KB gzip on both Vue versions.
  The two Vue 3 flags save about 3.7 KB more on Vue 3 (the Vue 3 saving minus the Vue 2 saving).
• A full page loads the same chunks as in feat: load pages faster by splitting the frontend into optional features #1235.
  In production builds, it loads about the same bytes: 3 KB less on Vue 3, 1 KB more on Vue 2.
  In development builds, it loads 1.9 to 3.1 KB more on both Vue versions.
• The dist folders have the same files as the feat: load pages faster by splitting the frontend into optional features #1235 build.
• I ran tests/unit/frontend*_test.py once on each of three environments (Vue 3 with ipywidgets 8, Vue 2 with ipywidgets 8, Vue 2 with ipywidgets 7): 84 passed on each.
  Two of these tests are new:
  • test_output_renderers_out_of_core checks for all 8 builds that a class name of the Markdown renderer is in the output-widget chunk and not in the core.
  • test_vue3_feature_flags_defined checks that the Vue 3 cores no longer name the two flags, and still name __VUE_PROD_DEVTOOLS__.
• I ran 6 browser tests of tests/integration/frontend_chunks_test.py once on each of the same three environments, with the flask and the starlette server: 12 passed on each.
  The tests are test_full_preloads_sync, test_full_display_outputs, test_minimal_lazy_controls, and three new ones:
  • test_minimal_lazy_output: a minimal page without an Output widget does not load output-widget; a page with one loads it once, and its Markdown outputs render.
  • test_full_output_widget_request_fails: the preloaded output-widget chunk fails once with a 404; the page still starts, loads the chunk on first use, and renders the outputs.
  • test_minimal_soft_remount_keeps_one_widget_renderer: after a soft remount, the stand-in holds one widget renderer, not two.
    I removed the fix once and rebuilt: this test then failed with 2 renderers.
    After I put the fix back, the rebuilt Vue 3 dist was byte for byte the same as before.

Gaps

• I did not measure load or run times on this branch.
  An earlier CPU profile of the earlier perf: let minimal pages without an Output widget skip the Jupyter output renderers #1236 builds gave estimates for the renderer code.
  It runs for about 2 ms at normal CPU speed and about 10 ms at 4x slowdown.
  The same profile found no clear time change from the two Vue 3 flags.
  I did not repeat that profile on this branch.
• I did not test third-party widgets that call the page's RenderMimeRegistry themselves.
  In minimal, such a widget gets an error for most calls until output-widget loads.
• The docs name only rendering calls, such as createRenderer, as calls that throw before output-widget loads.
  Other calls, such as getFactory or a read of mimeTypes, throw too.
• On Vue 3, Vue no longer sets the globals window.__VUE_OPTIONS_API__ and window.__VUE_PROD_HYDRATION_MISMATCH_DETAILS__ at startup, because the build now fixes their values.
  I know of no code that reads them.

Align results

The #1235 description lists the earlier questions of this work, Q0 to Q46.
These three settle this PR:

• Q41, reviewers: Opus agents do the work, and astra and gpt-6.1-sol do the reviews. Reviews run in parallel with builds, and do not rebuild.
• Q45, plain features in a stack: perf: let minimal pages without an Output widget skip the Jupyter output renderers #1236 becomes the bottom of a GitHub stack with plain features. jQuery and Lumino each get their own PR (the jquery PR (perf: let minimal pages without jQuery widgets skip loading jQuery (opt-in) #1241) and the lumino PR (perf: let minimal pages skip most of Lumino until a widget needs it #1240)): on in full (as in feat: load pages faster by splitting the frontend into optional features #1235), off in minimal, opt in with +jquery and +lumino. jupyter-controls and output-widget bring what their code imports. As a driver default, the create_view hook, the prototype swap and the guessing go. This PR keeps jQuery and Lumino in the core, as in feat: load pages faster by splitting the frontend into optional features #1235.
• Q47, jQuery opt-in and Lumino lazy: jQuery is opt-in with a clear error that names +jquery, and no lazy load. Lumino loads lazily with one warning that names the module that asked for it. Both changes are in their own PRs. The semver port is dropped. As a driver default, the stand-ins for @jupyterlab/coreutils, the kernelspec API and minimist are dropped too. The work file gives "under 2 KB, under 1 ms" as the reason. An earlier measurement put the minimist and kernelspec stand-ins at about 1.6 KB gzip and under 0.5 ms at normal CPU speed. Without the stand-ins, all of @jupyterlab/coreutils and @lumino/polling stay in the core. Nobody measured that total on these branches.

Crossreview results

Rebase onto #1235 at e8766df (2026-10-23): one new test: commit expects the vue-sfc warning that ipyvue 3.1 gives on minimal Vue 3 pages. astra and gpt-6.1-sol compared the old and new commits with git range-diff and found no regressions.

Round 1 had two reviewers: astra and gpt-6.1-sol.
Both come from one vendor, OpenAI. Maarten chose them in Q41.
Both reviewed the two commits of this PR (6a7caa63 and 6e151ed0) from the commits alone, without builds or tests.
Neither reviewer found a defect, so there are no CRITICAL, HIGH or MEDIUM findings, no fixes, and no small fixes.
Constitution change: None.

Round 2: astra and an Opus agent checked the PR descriptions against the code, and this version of the description has their corrections.

🤖 Generated with Claude Code