Version Packages

[hash-release]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@hashintel/ds-components@0.3.2

Patch Changes

• Extract the contour plot into a ContourSurface component shared by the sweep and optimization surfaces, decouple SweepNavigator from the experiments context, and add Storybook stories for each component's states. Sweep navigation is continuous: a point selection uses a single-thumb slider, slider moves commit during the drag, and charts keep their axes, grid and size while frames stream in. Slider treats 0 as a real value, and useElementSize follows an element that mounts after the first render. (@kube, #9421)

• Optimization studies with two or more optimized numeric parameters gain a Surface section: a contour of the objective over two chosen parameters, computed locally against the study's frozen model, with the study's trials as markers. Sliders and clicks move the selected point, which refines with escalating batches. Slider accepts step and onChangeEnd. (@kube, #9398)

• Add Breadcrumb component (@alex-e-leon, #9502)

@hashintel/petrinaut@0.0.20

Patch Changes

• Arc weight labels sit on the Adaptive Bezier curve instead of floating at the straight-line midpoint between the arc's endpoints. (@kube, #9354)

• The create-experiment drawer accepts an ad-hoc scenario when "No scenario" is selected: Initial State + Parameters defined inline compile through a generated, never-persisted scenario at experiment start. (@kube, #9288)

• Fix three defects in the ad-hoc scenario form: the optimize bounds popover
  ignored every press (Min, Max, Step and Scale were uneditable, and each press
  dismissed it), a focused section painted over the sticky header of the section
  hosting it, and the experiment drawer's computed initial state grew unbounded
  instead of scrolling in its own region. (@kube, #9473)

• The create-optimization drawer offers "Ad-hoc (define inline)": Initial State + Parameters with Optimize selections compile to a generated, never-persisted scenario whose generated parameters the optimization manifest binds to their optimize domains. (@kube, #9289)

• Quick simulation runs an ad-hoc scenario when none is selected: a "Define initial state" drawer in the simulation settings panel edits token counts and values that compile through a generated, never-persisted scenario. (@kube, #9287)

• Behind the new experimental "Ad-hoc scenarios" setting, the scenario creation form authors scenarios through the ad-hoc form: exposed Variables become the saved scenario's tunable parameters, and the definition persists as initialState.type: "adhoc". (@kube, #9368)

• Add the ad-hoc scenarios user-guide page and register it with the in-app AI assistant's doc reader. (@kube, #9290)

• The ad-hoc scenario form matches the ratified prototype: gutter-cycled row kinds (fixed → dynamic → count-optimized) with quiet count strips, shared columns with a wash and derived cells, a phantom trailing row, place totals, compact Variables blocks, and Monaco expression editors type-checked live through a new ad-hoc LSP session with diagnostics rendered on closed cells. (@kube)

• Transition kernels, lambdas, and differential equations can now be written as a bare function body ending in return, like metrics and scenario code: input (or tokens for dynamics) and parameters are in scope ambiently, with full editor type checking and completions. The export default TransitionKernel(...) / Lambda(...) / Dynamics(...) module form is still accepted, and the editor picks the right checking mode per form. Default templates, built-in examples, and the AI assistant now use the bare form. Visualizers are unchanged. (@kube, #9370)

• The canvas renders centered on the net from its first frame, instead of jumping there after a first paint at the origin. Component instances grow with their port count so their ports have room, and auto-layout on import no longer depends on the compact/classic setting. (@kube, #9352)

• Remember the canvas viewport per net, so switching between nets or reloading returns to the same position and zoom. (@kube, #9510)

• Added a command registry for host-rendered command palettes: createCommandRegistry() and combineCommandRegistries() in core, and CommandRegistryProvider, useCommand(command, { when }), useCommands(), and formatShortcutKeys() in @hashintel/petrinaut/react. The editor registers its undo/redo, tool, auto-layout, search, and panel commands. Petrinaut ships no palette. (@kube, #9457)

• The experiment and optimization result drawers keep the summary, navigator and surface still at the top; only the step and metric lists scroll, with the step table's header pinned. (@kube, #9422)

• Experiments can sweep scenario parameters. A Sweep toggle per numeric parameter defines the values, and the results drawer gains a navigator: only the selected combination computes, in escalating run batches that stream into the charts, and visited combinations keep their results. (@kube, #9093)

• The WebGPU backend matches the CPU engine's stochastic semantics and frame numbering, streams metric frames per dispatch chunk, and runs range sweeps from a per-run parameter buffer. It sizes and calibrates metric histograms and typed-place buffers from the model and device instead of refusing or warning, and tiles experiments larger than the device's limits with bit-identical results. Sweep batch instantiation no longer blocks the main thread: scenario compilation is prepared once per experiment and per-run values travel as one typed-array plan. (@kube, #9454)

• Distribution heatmaps in experiment metric charts render as a magma-coloured density image in one draw call, so they stay fast at high run counts. (@kube, #9450)

• Swept parameters declare an interval instead of a value count, and the navigator becomes a range slider per parameter: the whole interval by default, resizable, collapsible to a point. A range selection samples the region and streams the merged distribution over it. The slider is a two-thumb range control local to the navigator, whose thumbs may coincide. (@kube, #9396)

• Extract the contour plot into a ContourSurface component shared by the sweep and optimization surfaces, decouple SweepNavigator from the experiments context, and add Storybook stories for each component's states. Sweep navigation is continuous: a point selection uses a single-thumb slider, slider moves commit during the drag, and charts keep their axes, grid and size while frames stream in. Slider treats 0 as a real value, and useElementSize follows an element that mounts after the first render. (@kube, #9421)

• Optimization studies with two or more optimized numeric parameters gain a Surface section: a contour of the objective over two chosen parameters, computed locally against the study's frozen model, with the study's trials as markers. Sliders and clicks move the selected point, which refines with escalating batches. Slider accepts step and onChangeEnd. (@kube, #9398)

• Run an experiment's runs in parallel across several Web Workers. (@kube, #9162)

  An experiment used to run every one of its runs in a single worker, using one core however many the machine had. Runs are independent, so they now split across one worker per logical core (minus one, so the editor stays responsive), capped at the run count. Measured at ~4x on 8 shards on a 10-core machine.

  Sharding cannot change what an experiment reports. Per-run seeds derive from the run's global index rather than its position within a shard, so run i gets the same seed whichever worker owns it, and each worker's per-frame statistics recombine through the metric accumulator monoids (empty/merge). Output is byte-identical at every shard count while every shard still has an active run. A frame is only finalised once every still-running shard has reported it, with finished shards dropped from that watermark rather than blocking it — so once a whole shard's runs have ended early (for example by deadlock), that shard's completed runs stop contributing samples to later frames, where a single simulator would keep sampling their frozen state.

  Scalar metric frames now carry their pre-reduction accumulator state, because frameValue is already reduced and a mean of means is not a mean.

  Hosts can cap or pin parallelism with experimentShardCount on ExperimentsProvider, or shardCount on createMonteCarloExperiment.

• Add a navigation prop to Petrinaut: a router-neutral controller through
  which the host can read and drive the app location (mode, Simulate section and
  resource, scenario, subnet, selection, and creation drawers), making them real
  browser history destinations. A creation drawer now layers over the record
  already open instead of closing it, and the hamburger menu hides Layout on
  a read-only net. (@kube, #9426)

• Add an optional per-place token capacity. (@kube, #9177)

  A place can now declare a maximum number of tokens it will hold, set from the place properties panel. Useful for supply-chain style models with finite storage. It also converts frames from growable to fixed-size, which is the precondition for a fixed-layout GPU or WASM path.

  Capacity participates in transition enablement, following the standard Petri-net capacity constraint: a transition cannot fire if doing so would take any output place above its capacity. Output tokens are applied at the end of a frame, so the check accounts for what transitions earlier in the same frame have already committed. Several transitions feeding one capped place cannot collectively overflow it.

  Deadlock detection includes the same check, so a net whose only remaining transitions are blocked by full output places is reported as deadlocked rather than stepping to maxTime with nothing happening.

  Nets without capacities are unaffected: the constraint tables are empty and the hot path skips them.

• Export the EditorGlobalMode and SimulateViewMode types, so a host encoding Petrinaut's navigation state into its own router can spell both vocabularies and fail its build when either gains a member. (@kube, #9491)

• Reject net identifiers that collide with Object.prototype member names (__proto__, constructor, ...) at file import and before simulation, and store user-authored keys in prototype-free records. Place visualizer code now runs under the same sandbox hardening as scenario code. (@kube, #9222)

• Scenario code (parameter overrides, per-place expressions, code-mode initial state) now compiles through the HIR and runs in an interpreter: no new Function, no sandbox. Out-of-subset code fails with positioned errors in the editor and at compile. Stricter than the old evaluator: == is strict, conditions and &&/|| take booleans, unknown place names in code mode are errors, and array methods beyond .map/.reduce/.concat are unavailable. Array.from({ length: n }, fn) desugars to range(n).map(fn). (@kube, #9332)

• Range sweeps rotate each axis's low-discrepancy parameter draws by a seed-derived shift, so experiments with different seeds explore different value sequences. (@kube, #9435)

• Sidebar list labels use the full row width and truncate with an ellipsis consistently; the row menu button only takes space while hovering the row or while its menu is open. (@kube, #9377)

• Add generic host-rendered AI composer controls and a persistent inline Voice session, protected
  active conversations, keyboard fallback, and one-answer buffering while the normal chat stream
  settles. Add the provider-neutral renderVoiceMode contract and export PetrinautAiInputMode, with
  the selected input mode and mode-change callback available to host-rendered Voice modes. Include
  stable finalized-text submission, conversation identity, stop handling, schema-validated
  interactive-tool text mapping, explicit separate-message targeting for corrections, and a
  queue-aware voice submission path. Present text and voice through one transcript and composer whose
  trailing action switches between waveform, Send, and Stop. renderComposerControl remains a
  supported public seam for hosts that only need their own control beside the message box,
  independently of Voice mode. Surface assistant request failures as error toasts instead of
  transcript entries. (@kostandinang, #9355)

  Render every live Voice surface from a session snapshot the host reports through
  reportVoiceSessionState, so hosts describe their session while Petrinaut owns its chrome. Replace
  the composer with a low-profile Voice dock -- a canvas ribbon of three travelling waves, one leading
  and two trailing it more faintly, fading out at both ends, opening with the microphone while
  listening and moving under its own power through the same drawing path while the assistant speaks,
  near flat while neither holds the turn, and naming one phase at a time -- with an announced phase and
  reduced-motion behavior. Curve the raw microphone level before it drives the waves, so an ordinary
  speaking level is plainly visible rather than a flicker above the line and a raised voice still has
  somewhere to go. Sample the level per animation frame rather than through React, so drawing costs no
  re-renders, and ease it against the frame delta so it looks the same at any refresh rate. Hold spoken
  turns out of the transcript until the session ends, then reveal them together under a turn-count
  divider, while typed messages and interactive tools awaiting an answer stay visible throughout. Let a
  per-session Show transcription in chat action write those turns into the conversation as they land
  instead. Keep every session control -- transcription, the microphone toggle, Resume, Reconnect, and
  End -- in the dock, leaving the canvas toolbar untouched. Add setMicrophoneMuted to the Voice mode
  controls and a muted session phase, so muting stops capture without interrupting what the assistant
  is saying, unlike pausing. Surface voice recovery failures as toasts with privacy-safe diagnostic
  references, and request one-time consent before the host starts the microphone. Mark persisted spoken
  messages and the exact interactive-tool answer completed by Voice with an inline Voice chip ahead of
  the words themselves.

  End Voice mode before submitting typed text exactly once through the shared composer, preserving the
  draft if handoff fails. Pause active media before the AI panel closes and reopen the mounted session
  paused. Provisional transcription and Realtime audio remain ephemeral rather than becoming
  persisted chat history.

• Storybook gains a "With real optimizer" story that runs optimization studies against a local Petrinaut Optimizer service; start it with turbo run dev --filter @hashintel/petrinaut -- --with-optimizer-service. (@kube, #9395)

• A sweep's range selection runs as one stochastic simulation over the ranges: every run draws its own value per ranged parameter, and the metric distribution over the region streams live. ExperimentRequest carries optional per-run overrides (runs), forwarded by the worker-pool backend and refused by the WebGPU backend. (@kube, #9419)

• Sweep compute is reused and pipelined: simulation workers persist across batches, surface cells sample in batched chunks in quad-tree order across the CPU pool, the navigator's selection streams first while the ladder pipelines its rungs, and one GPU backend (device, shader, calibration) with a small preview tile serves a session's batches. The surface reads a cell's value from the last sampled frame, so terminating nets fill it; charts mount their axes when the drawer opens and apply data once per animation frame. (@kube, #9474)

• The frame inspector draws a frame's distribution as a canvas histogram with value and count axes. Distribution frames carry their bins' extent and heatmaps paint each bin across the rows it covers, so mixed strides no longer stripe, and streamed updates ease in instead of snapping. The sweep surface navigates by drag as well as click and marks the navigator's position, and the summary lists every batch computing in parallel. Parameter sweeps and the optimization surface are experimental settings, off by default. (@kube, #9478)

• Sweeps with two or more swept parameters gain a Surface section: a contour of a metric's final value over two chosen parameters, filled in live on a background lane. Clicking it moves the navigator to the nearest combination. (@kube, #9369)

• Add an experimental WebGPU compute backend for experiments, chosen per experiment behind a user setting. It runs the net's lowered HIR on the device, declines nets it cannot run so they fall back to the CPU, and agrees with the CPU in distribution rather than seed for seed. A Compilation panel, also behind a setting, shows what the compiler made of each condition, kernel and equation. (@kube, #9179)

• Add the worksheet keyboard-flow primitives (FocusStack, FocusRoot, useFocusGrid, useFocusStops): reusable arrow-key routing, per-group focus memory, and roving tab stops for panels composed of grids and controls. (@kube, #9411)

• Nets and optimization manifests can be imported and exported in YAML and JSON. (@kube, #9379)

• Updated dependencies:

  • @hashintel/petrinaut-core@0.0.5
  • @hashintel/ds-components@0.3.2

@hashintel/petrinaut-core@0.0.5

Patch Changes

• Behind the new experimental "Ad-hoc scenarios" setting, the scenario creation form authors scenarios through the ad-hoc form: exposed Variables become the saved scenario's tunable parameters, and the definition persists as initialState.type: "adhoc". (@kube, #9368)

• Add the ad-hoc scenarios user-guide page and register it with the in-app AI assistant's doc reader. (@kube, #9290)

• Ad-hoc scenario synthesis: form state, deterministic adhoc.* parameter names, synthesis to a generated code-mode scenario plus optimized fields with typed domains, the transform to optimization-manifest bindings, row-kind cycling and shared-column transitions, place totals, and an LSP session kind that type-checks every value expression. (@kube, #9284)

• Transition kernels, lambdas, and differential equations can now be written as a bare function body ending in return, like metrics and scenario code: input (or tokens for dynamics) and parameters are in scope ambiently, with full editor type checking and completions. The export default TransitionKernel(...) / Lambda(...) / Dynamics(...) module form is still accepted, and the editor picks the right checking mode per form. Default templates, built-in examples, and the AI assistant now use the bare form. Visualizers are unchanged. (@kube, #9370)

• Added a command registry for host-rendered command palettes: createCommandRegistry() and combineCommandRegistries() in core, and CommandRegistryProvider, useCommand(command, { when }), useCommands(), and formatShortcutKeys() in @hashintel/petrinaut/react. The editor registers its undo/redo, tool, auto-layout, search, and panel commands. Petrinaut ships no palette. (@kube, #9457)

• The layout module exports the canvas geometry: render node dimensions (compactNodeDimensions, classicNodeDimensions, getComponentInstanceHeight), net bounds (getBoundsOfCenteredBoxes) and zoom limits (getMinZoomForBounds, ZOOM_PADDING). layoutNodeDimensions is now derived from the render dimensions instead of maintained by hand. (@kube, #9352)

• Add the @hashintel/petrinaut-core/experiments entry point: an ExperimentBackend interface, a worker-pool implementation of it, and selectExperimentBackend, which walks backends in preference order and records every refusal. (@kube, #9178)

• createJsonDocHandle accepts SDCPNInput, a loose authoring variant of SDCPN: extension fields may be omitted and are filled with plain-net defaults by the new normalizeSDCPN export. SDCPNInput and its member types are exported from the package root. (@kube, #8926)

• The WebGPU backend matches the CPU engine's stochastic semantics and frame numbering, streams metric frames per dispatch chunk, and runs range sweeps from a per-run parameter buffer. It sizes and calibrates metric histograms and typed-place buffers from the model and device instead of refusing or warning, and tiles experiments larger than the device's limits with bit-identical results. Sweep batch instantiation no longer blocks the main thread: scenario compilation is prepared once per experiment and per-run values travel as one typed-array plan. (@kube, #9454)

• Weighted-arc token combinations enumerate lazily in the same lexicographic order, so a transition with a weight-2 coloured input arc no longer materialises every combination per frame. Trajectories are unchanged for every seed. (@kube, #9386)

• Run an experiment's runs in parallel across several Web Workers. (@kube, #9162)

  An experiment used to run every one of its runs in a single worker, using one core however many the machine had. Runs are independent, so they now split across one worker per logical core (minus one, so the editor stays responsive), capped at the run count. Measured at ~4x on 8 shards on a 10-core machine.

  Sharding cannot change what an experiment reports. Per-run seeds derive from the run's global index rather than its position within a shard, so run i gets the same seed whichever worker owns it, and each worker's per-frame statistics recombine through the metric accumulator monoids (empty/merge). Output is byte-identical at every shard count while every shard still has an active run. A frame is only finalised once every still-running shard has reported it, with finished shards dropped from that watermark rather than blocking it — so once a whole shard's runs have ended early (for example by deadlock), that shard's completed runs stop contributing samples to later frames, where a single simulator would keep sampling their frozen state.

  Scalar metric frames now carry their pre-reduction accumulator state, because frameValue is already reduced and a mean of means is not a mean.

  Hosts can cap or pin parallelism with experimentShardCount on ExperimentsProvider, or shardCount on createMonteCarloExperiment.

• Expose the selection vocabulary as data: selectionItemTypes and
  canonicalizeSelection are available from a dependency-free
  @hashintel/petrinaut-core/selection entry, so hosts can validate and order
  selection coming from a URL or an HTTP request without pulling the model or any
  React code. (@kube, #9426)

• Add an optional per-place token capacity. (@kube, #9177)

  A place can now declare a maximum number of tokens it will hold, set from the place properties panel. Useful for supply-chain style models with finite storage. It also converts frames from growable to fixed-size, which is the precondition for a fixed-layout GPU or WASM path.

  Capacity participates in transition enablement, following the standard Petri-net capacity constraint: a transition cannot fire if doing so would take any output place above its capacity. Output tokens are applied at the end of a frame, so the check accounts for what transitions earlier in the same frame have already committed. Several transitions feeding one capped place cannot collectively overflow it.

  Deadlock detection includes the same check, so a net whose only remaining transitions are blocked by full output places is reported as deadlocked rather than stepping to maxTime with nothing happening.

  Nets without capacities are unaffected: the constraint tables are empty and the hot path skips them.

• Reject net identifiers that collide with Object.prototype member names (__proto__, constructor, ...) at file import and before simulation, and store user-authored keys in prototype-free records. Place visualizer code now runs under the same sandbox hardening as scenario code. (@kube, #9222)

• Scenario code (parameter overrides, per-place expressions, code-mode initial state) now compiles through the HIR and runs in an interpreter: no new Function, no sandbox. Out-of-subset code fails with positioned errors in the editor and at compile. Stricter than the old evaluator: == is strict, conditions and &&/|| take booleans, unknown place names in code mode are errors, and array methods beyond .map/.reduce/.concat are unavailable. Array.from({ length: n }, fn) desugars to range(n).map(fn). (@kube, #9332)

• A sweep's range selection runs as one stochastic simulation over the ranges: every run draws its own value per ranged parameter, and the metric distribution over the region streams live. ExperimentRequest carries optional per-run overrides (runs), forwarded by the worker-pool backend and refused by the WebGPU backend. (@kube, #9419)

• Sweep compute is reused and pipelined: simulation workers persist across batches, surface cells sample in batched chunks in quad-tree order across the CPU pool, the navigator's selection streams first while the ladder pipelines its rungs, and one GPU backend (device, shader, calibration) with a small preview tile serves a session's batches. The surface reads a cell's value from the last sampled frame, so terminating nets fill it; charts mount their axes when the drawer opens and apply data once per animation frame. (@kube, #9474)

• The frame inspector draws a frame's distribution as a canvas histogram with value and count axes. Distribution frames carry their bins' extent and heatmaps paint each bin across the rows it covers, so mixed strides no longer stripe, and streamed updates ease in instead of snapping. The sweep surface navigates by drag as well as click and marks the navigator's position, and the summary lists every batch computing in parallel. Parameter sweeps and the optimization surface are experimental settings, off by default. (@kube, #9478)

• Transitions whose lambda reads no input tokens skip combination enumeration and evaluate the lambda once against the first tokens in place order. Trajectories are unchanged for every seed. (@kube, #9389)

• Optimization trials run their seeded replicates in parallel as one sharded experiment. The Monte Carlo worker protocol attaches to any thread runtime, and the CLI's --threads <n> bounds the workers, defaulting to one per core minus one. (@kube, #9358)

• Add an experimental WebGPU compute backend for experiments, chosen per experiment behind a user setting. It runs the net's lowered HIR on the device, declines nets it cannot run so they fall back to the CPU, and agrees with the CPU in distribution rather than seed for seed. A Compilation panel, also behind a setting, shows what the compiler made of each condition, kernel and equation. (@kube, #9179)

• Nets and optimization manifests can be imported and exported in YAML and JSON. (@kube, #9379)

@hashintel/brunch-agent-plugin-sdcpn@0.0.0

Patch Changes

• Updated dependencies:
  • @hashintel/petrinaut-core@0.0.5

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 4, 2026 3:57pm UTC
hashdotdesign-tokens	Ready	Preview	Sep 4, 2026 3:57pm UTC
petrinaut	Ready	Preview	Sep 4, 2026 3:57pm UTC
petrinaut-docs	Ready	Preview	Sep 4, 2026 3:57pm UTC

[cursor]

PR Summary

Medium Risk
Merging triggers npm publish of petrinaut-core and petrinaut with simulation, experiment sharding, stricter scenario semantics, and security hardening already described in the release notes—not just metadata edits.

Overview
This Version Packages changesets release clears the pending .changeset entries and publishes new semver versions with consolidated CHANGELOG notes—no application source changes in the diff itself.

@hashintel/petrinaut@0.0.20 and @hashintel/petrinaut-core@0.0.5 ship the accumulated Petrinaut work: ad-hoc scenario authoring and inline scenarios for simulation/experiments/optimization; parameter sweeps with range sliders, contour surfaces, pipelined compute, and optional WebGPU experiment backend; parallel Monte Carlo runs across workers with monoid-safe metric merging; HIR-backed scenario execution (stricter than the old evaluator); place token capacity; host navigation and selection vocabulary exports; command registry hooks; AI voice composer integration; YAML/JSON import-export; and assorted canvas and UX fixes.

@hashintel/ds-components@0.3.2 adds Breadcrumb, shared ContourSurface / sweep navigator improvements, and optimization surface slider APIs. @hashintel/brunch-agent-plugin-sdcpn is bumped to 0.0.0 with a dependency-only changelog entry on petrinaut-core@0.0.5.

^{Reviewed by Cursor Bugbot for commit 1678143. Bugbot is set up for automated code reviews on this repo. Configure here.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 60.89%. Comparing base (eea4e1f) to head (d78f871).
⚠️ Report is 7 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #9340   +/-   ##
=======================================
  Coverage   60.89%   60.89%           
=======================================
  Files        1461     1461           
  Lines      146725   146725           
  Branches     6744     6744           
=======================================
  Hits        89343    89343           
  Misses      56264    56264           
  Partials     1118     1118           

Flag	Coverage Δ
apps.hash-api	15.41% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.