FE-1583: Run the connected optimizer's trials through the experiments backend

[kube]

Important

Experimental
Behind the In-browser optimization feature flag.

Summary

Before this PR, PetrinautOptimizationContext accepted a self-contained remote capability, and a study's trials were numbers in a table. Study drawer's Surface could sample the objective locally, but it knew nothing of where the optimizer was looking, and the optimizer could not use the browser's compute at all.

This PR lets the context carry a connected optimization source and runs its trials through the experiments layer. A new runDetachedObjective action runs a frozen study's objective at one parameter point on the chosen backend with pinned seeds and streams its frames. Optimizations provider evaluates each trial through it, follows the trial in the navigator, streams its metrics live, and after the study settles refines the best point, or whichever point the user picks. Study drawer shows the optimizer's own steps as dots on the Surface with the field filling in as steps report, a Stop that keeps the study and a Continue that adds steps to it, up to four steps in flight with an activity list, and one metric line always in view. Form gains Runs per step, Parallel steps and the experiments' Backend switch. A remote capability such as the HASH host's NodeAPI bridge is unaffected, and the connected source only takes effect with the experimental setting on.

9516.mp4

Links

• Ticket FE-1583

Changes

Feature flag

• New experimental setting In-browser optimization under Settings > Simulation

  Row shows only when the host supplies a connected optimizer.

• useOptimizationSource is the one place that reads the context and the setting

  A remote capability passes through as is.
  A connected source passes through only with the setting on, else the tab stays hidden and nothing connects; turning it off cancels connected runs and disposes.

Experiments layer

• runDetachedObjective runs a study's objective at one point and streams it

  Compiles once per study, chooses the backend once per study and backend with the experiments' registrations at a third of the CPU pool, pins per-run seeds on the CPU pool only, streams frames and progress.
  Settles with per-run results or a reason: compile diagnostics, backend refusals, failed runs, cancellation.

• Runs queue per queueKey while sharing one compiled snapshot per cacheKey

  Parallel steps of one study overlap on the backend; refinement rungs of one point stay in order.

• A cancelled run settles instead of hanging its queue

  Instantiation gets its own abort controller, mirrored from the run signal only until the backend handle exists.
  A later cancel goes through the handle, whose cancelled event releases the queue to the next run.

• Sweep throttle moves to experiments/shared for reuse

Optimization channel and provider

• Channel evaluates a trial as one detached objective run with the trial's seeds and the study's backend

  Objective is the mean of the per-run finals when the backend reports them, else the last frames' value; a failure prunes the trial with its reason; the channel never throws.

• Records carry computeBackend, its fallback reason, the study's axes, parallelism, resumable, a navigation, a selection stream, inFlight steps and an activity list

  Navigation follows the most recently started step; a user move stops following.
  Selection stream carries the followed step's frames, then the refined point's.

• Study settles on its best point when it completes or stops

  Navigation moves to the best step while it still follows, and refinement starts there.

• Point refinement climbs the run ladder to 100 runs for the navigated point

  Starts when the study settles or the user moves, cancels and restarts on a navigation change, caches per point, stops with the reason on failure.
  Between rungs it compares the point's estimate with the best: a point that cannot beat the best by 2.5 standard errors stops after its first rung with the note "8 runs · cannot beat the best"; the best point refines fully.

• extendOptimization(id, steps) continues a finished or stopped connected study

  Provider calls extendOptimizationRun, re-attaches from the record's cursor, turns following back on; the steps continue Optuna's numbering.
  Stop keeps the attachment so the segment's terminal event lands and marks the record resumable.

• Removing a connected record releases its study in the worker
• Connected runs are not written to session storage since they cannot outlive the page

Form and drawer

• Create form gains Runs per step, Parallel steps, Seed and the Backend switch

  Runs per step lands in execution.seedsPerTrial; the total-steps rule multiplies by it.
  Parallel steps is 1 to 4 and shows for a connected source only.
  Seed starts at a fresh random value each time the form opens and seeds both the proposals and the runs; every study used to share one fixed seed, so two studies drew the same random steps.
  Backend cell is the experiments' shared toggle, shown for a connected source with WebGPU on; it stays disabled with the reason because the GPU backend cannot compute an expression objective.

• Connected study drawer lays everything out to fit one screen

  A summary strip carries status, finished over requested steps, best value, the backend badge, the steps bar and the "N computing" chip.
  Parameters band holds a slider per axis, Follow steps and the refinement status.
  Surface sits beside the objective chart, the Steps table fills the rest, and the footer holds Stop, or Remove and Continue with a step count.

• Surface shows the optimizer's steps only

  Each step with an objective is one sample of the field at its projected position and a dot, the best emphasized, pruned steps hollow.
  A step in flight enters the field with its running value, and so does the refined point after settle, so the plot fills in as the study streams.
  While a running study is followed the plot and the sliders are display-only; once the study is terminal, or Follow steps is off, a click or drag commits the navigation.

• Remote study drawer keeps its locally sampled Surface and its Cancel
• Shared pieces move to SimulateView/shared: useGpuAvailability, ComputeBackendToggle, ComputeBackendBadge, MetricTiles, ComputeActivity, SummaryStrip
• Slider in @hashintel/ds-components accepts disabled

Stories and example

• Storybook Simulate / Browser optimizer (real) runs the real Pyodide optimizer against the real experiments backends

  SirCpu, SirGpuRequested, SupplyChain, VaccinationCampaign start a study on load; Manual opens the create form.
  Shared harness simulate-view-story-harness.tsx with AutoStudy; the synthetic-optimizer story is renamed RunSupplyChainOptimizationSyntheticOptimizer.

• New core example Vaccination Campaign built for optimization demos

  SIR with pre-wave vaccination coverage and contact reduction, a Total cost objective whose minimum lies inside the domain, GPU-eligible; listed under Load example.

Known issues

• GPU studies are unreachable from the form until the WebGPU backend computes expression metrics

  Record, channel and registrations already carry the choice.

Test coverage

• provider/detached-objective.test.ts:

  Streaming, pinned seeds on the CPU and none on the GPU, fallback re-pinning, failure reasons for refusal, errored runs and compile errors, one walk per study and backend, per-key queueing, a cancelled batch settling and releasing the queue, distinct queue keys side by side, cancel and dispose.

• optimizations/provider.test.tsx, provider/connected-study.test.ts, provider/point-refinement.test.ts, provider/point-refinement/objective-estimate.test.ts, provider/activity-registry.test.ts, channel/create-optimization-channel.test.ts:

  Setting gate on, off and toggled, remote unaffected, trials through runDetachedObjective with seeds and backend, following and user moves, settle on the best, the ladder and its early stop, restart on navigation change, stop then continue from the cursor, refused continuation, removal releasing the study, activity throttling.

• create-optimization-drawer.test.tsx, view-optimization-drawer.test.tsx, study-summary-strip.test.ts, steps-table.test.ts, optimization-navigator.test.tsx, optimization-surface/surface-plot.test.tsx, simulate-view.test.tsx, viewport-settings-dialog.test.tsx:

  Runs per step, parallel steps, the typed seed and its limit, the backend option, Stop, Stopped and Continue, progress bars and the computing chip, the note on a refused point, trials-only field and provisional samples, display-only while following, tab gating, default off.

• examples/vaccination-campaign.test.ts, webgpu/compilation-report.test.ts:

  GPU eligibility and readiness of the example, its default marking, and a seeded check that the cost at the valley is below the domain's boundary points.

• Storybook DOM probes and screenshots:

  SirCpu streaming, Stop at step 2 then Continue +4 to Complete 6 / 6, a hopeless pick stopping after 8 runs, the best point climbing to 100 runs.

How to test

• yarn workspace @hashintel/petrinaut dev
• Simulate > Browser optimizer (real) > SirCpu

  Expect the runtime to download once, then steps as dots on the Surface, the field filling in, the metric line streaming

• Stop

  Expect Stopped, the navigation on the best step, refinement to 100 runs

• Continue

  Expect steps resuming with continued numbering, then Complete

• Click a corner of the Surface

  Expect "0 of 8 runs — refining" then "8 runs · cannot beat the best"

• Click next to the best dot

  Expect refinement climbing to 100 runs and the metric line restreaming

• Simulate > Browser optimizer (real) > VaccinationCampaign

  Expect a valley on the Surface with the best dot on its floor

• Simulate > ViewOptimizationDrawer stories

  Expect the running, complete, stopped, fallback and remote variants to render

[vercel]

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

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 4, 2026 12:39pm UTC
petrinaut	Ready	Preview	Sep 4, 2026 12:39pm UTC
petrinaut-docs	Ready	Preview	Sep 4, 2026 12:39pm UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 4, 2026 12:39pm UTC