FE-1690: Toggle optimization from the experiment creation form

[kube]

Important

Experimental
Behind the Parameter sweeps and In-browser optimization feature flags.

Summary

Before this PR, optimizing a sweep took two steps. The Create Experiment drawer offered a Sweep toggle on each numeric value and created an idle sweep; the optimizer started later from a purple Optimize button on the sweep's Parameters card, which asked for the metric, the direction and the step count in a popover, and a sweep could be optimized again and again into one strip.

This PR starts the study with the experiment. Where the in-browser optimizer can drive the sweep, the interval toggle reads Optimize instead of Sweep; the first toggle adds an Objective section to the drawer, metric over the experiment's metric drafts, Maximize or Minimize, and the number of steps, with a helper line that pre-checks the optimizer's step budgets. The footer reads Optimize, and submit is one gesture: create the experiment, start the study, then select it, so the results drawer opens already driving with the Parameters card purple, the headline counting steps and the cards and table mounted. A study that cannot start removes the experiment again and names the reason in the footer. The Parameters card keeps only Stop while a study drives; one experiment carries one study, and a new search is a new experiment. Without the optimizer the toggle still reads Sweep and the sweep opens idle as before.

9689.mp4

Links

• Ticket FE-1690
• Docs Sweep orchestration
• Docs The ad-hoc scenario form
• Docs Experiments › Optimizing a sweep
• Docs Ad-hoc Scenarios › Interval selections

Changes

Create Experiment drawer

• The interval toggle reads Optimize where the optimizer can drive the sweep

  Parameter sweeps and In-browser optimization on with an optimizer that runs in the browser gives Optimize; Parameter sweeps alone, or a remote-only optimizer, keeps Sweep. The word never depends on how many toggles are on.
  Min and Max cells are the whole interval editor; the Step and Scale cells and the boolean two-way choice are deleted.

• Objective section

  Appears with the first Optimize toggle, between Scenario and Constraints. Metric over the drawer's metric drafts, Maximize or Minimize, Steps with a default of 30.
  A reserved helper line reads the plan or names the budget the optimizer would refuse, mirrored from the core's step limits.
  While an ad-hoc bound fails to synthesize the section stays mounted, the footer keeps Optimize disabled and the summary line carries the error, so the drawer does not change shape mid-edit.

• Submit creates, starts and selects in that order

  createExperiment resolves the record and no longer selects it; the drawer starts the study through createOptimization and selects afterwards, so the results drawer mounts with the study in place.
  A failed start removes the experiment, keeps every field and puts the reason in the footer.

• Ad-hoc experiments optimize too

  The manifest is built from the record's compiled scenario; every generated parameter is an axis. Constraints stay saved-scenario only.

Results drawer

• Stop is the only card control

  The Optimize button, its popover and the multi-study strip machinery are deleted; the strip and the cards read one study.

• The drawer opens in the driving state for an Optimize-created experiment

  Every study slot renders at initializing, so the shape never changes afterwards; a Sweep-created experiment opens idle as before.

Public API

• createExperiment resolves with the ExperimentRecord, which carries the scenario it compiled

  ExperimentRecord.scenario is a new required field; callers select the experiment themselves. Patch changeset for @hashintel/petrinaut.

Known issues

• No Optimize again on an experiment

  A stopped or finished study cannot be restarted; a new search is a new experiment. Bringing it back needs the objective on the record and the multi-study strip.

• Sweep-only exploration with both flags on

  Press Stop right after the drawer opens, or turn In-browser optimization off to get the Sweep word back.

• The Surface pickers stay read-only while the study drives

  Unchanged behaviour, more visible now that every Optimize-created experiment opens driving.

Test coverage

• create-experiment.test.ts, provider.test.tsx, context.test.ts:

  The record carries its compiled scenario, creation resolves the record without selecting it, no await sits before the sweep session starts.

• ad-hoc-scenario-form.test.tsx, experiment-scenario-run.test.tsx:

  Interval selection under both words, numeric values only, Min and Max as the whole editor, gutter labels following the word.

• sweep-objective.test.ts, objective-section.test.tsx, create-optimized-experiment.test.tsx, sweep-optimizer.test.ts:

  Metric resolution and fallback, the step budgets, create then start then select with rollback, the manifest for saved and ad-hoc records.

• create-experiment-drawer.test.tsx, ad-hoc-interval-toggles.test.ts:

  The word per settings and source, the Objective section's gating and defaults, the footer label and budget error, submit reaching the fake optimizer, a failed start keeping the drawer open, a failing ad-hoc bound keeping the section and the word.

• experiment-results.test.tsx, view-experiment-drawer.test.tsx, sweep-objective-history.test.ts:

  Stop only while driving, the layout signature across study states, the single-study strip.

How to test

• Open Petrinaut preview on Vercel
• Viewport controls > Settings > Simulation > Parameter sweeps on, In-browser optimization on
• Menu > Load example > SIR Epidemic Model
• Switch to Simulate
• Experiments > Create
  • Scenario > Seasonal Flu
  • Flip Optimize on infected_ratio

    Expect an Objective section with Metric, Direction and Steps, and the Constraints section under it

  • Add a metric on Infected, choose Maximize, Steps 10

    Expect the footer to read Optimize

  • Steps 1001

    Expect the footer disabled with the budget message

  • Steps 10, then Optimize
• Results drawer

  Expect the Parameters card purple, the headline counting steps, Stop on the card, the Objective by step strip, the cards and the steps table mounted from the first frame

• Stop

  Expect the sliders unlocked, the outcome on the headline, no Optimize button

• Viewport controls > Settings > Simulation > In-browser optimization off
• Experiments > Create > flip the toggle on infected_ratio

  Expect it to read Sweep, no Objective section, footer Create sweep

[vercel]

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

Project	Deployment	Actions	Updated
hash	Building	Preview	Sep 13, 2026 2:04pm UTC
petrinaut	Ready	Preview	Sep 13, 2026 2:04pm UTC
petrinaut-docs	Ready	Preview	Sep 13, 2026 2:04pm UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 13, 2026 2:04pm UTC

[cursor]

PR Summary

Medium Risk
Touches the public createExperiment contract and experiment/optimization lifecycle orchestration; behavior is behind experimental feature flags but mis-ordering creation vs sweep session registration could break optimizer trials.

Overview
Moves optimization into the Create Experiment flow instead of starting it later from the sweep’s Parameters card. When in-browser optimization is enabled, numeric interval toggles read Optimize (otherwise Sweep); the first Optimize toggle adds an Objective section (metric, direction, steps) and constraints gating shifts to Optimize toggles. Submit creates the experiment, starts the study, then the drawer selects it so results open already optimizing—with Stop on the Parameters card and study UI (Objective by step, Steps columns, constraint/sensitivity cards) fixed at creation for optimize runs.

Public API change: createExperiment now resolves with the full ExperimentRecord (including compiled scenario) and no longer auto-selects the experiment; callers own navigation. The default context rejects if used outside ExperimentsProvider.

Ad-hoc / interval editing is unified: interval mode is Min and Max only (Step, Scale, and boolean optimize UI removed); gutter labels follow Sweep vs Optimize (Swept count / Optimized count).

Docs, Storybook notes, and tests are updated for the new creation contract, scenario on the record, and interval-selection behavior.

^{Reviewed by Cursor Bugbot for commit a533a06. Bugbot is set up for automated code reviews on this repo. Configure here.}