FE-1484: Run experiments from Petrinaut AI chat

[kube]

Important

Experimental optimization
Optimization requires Parameter sweeps and In-browser optimization. Simulation and the diagnostics API work without these flags.

Summary

Petrinaut AI can create an Experiment from chat, show its progress, and receive a captured result when it finishes. The same request can optimize a parameter range, then finish the requested runs at the best point.

The browser host owns validation and execution. Integrating applications consume a shared contract; Brunch continues to own its tool selection and agent behavior. The website demo uses scripted chat replies and real browser computation, with that boundary stated on screen.

Links

• Ticket FE-1484 (internal)
• Docs AI-created experiments and D2 diagrams
• Docs Client integration responsibilities
• Docs AI assistant › Experiments from chat

Blocked by

• Optimize a parameter sweep from its Parameters card #9650 provides experiment optimization

Changes

Contract and browser host

• Add shared Zod request/result schemas and the PetrinautExperimentHost interface

  Requests use saved scenario and metric IDs, fixed values or parameter ranges, and bounded run budgets. The host captures the model and validates references and code before creating an experiment.

• Protect computation inputs until the host captures the terminal result

  The user can inspect and cancel a running request. Later exploration preserves the result already returned to chat.

• Retain metric distributions for heatmaps and histograms, including after optimization

  Chat receives the mean of the final sampled distribution; the experiment keeps the spread across runs for exploration.

• Replace UI diagnostics polling with a correlated language-worker snapshot request

Chat and integration demo

• Show compact blue simulation cards and purple optimization cards with an active glow

  Status and run count sit beside the title; metrics share a row with the View experiment action. Progress animates between reported values. Active indicators stop on completion and respect reduced-motion preferences.

• Keep chat mounted across editor views and show completion notifications
• Add the /ai-experiments demo with planning turns, streamed replies, and completion figures from the returned experiment result

  The illustrative SIR model uses fixed simulation and search recipes. The demo documents separate client and host responsibilities; Brunch's production catalog, prompts, and transport remain unchanged.

• Externalize React runtime subpaths so the website can load the library in a browser

Known issues

• Execution lasts for the mounted browser host; durable jobs remain a follow-up
• Requests use saved scenarios and metrics on the CPU backend; metrics use the last sampled frame and means across runs
• Diagnostics requests have no timeout for an unresponsive language worker

The architecture guide marks these limits and the future host-orchestration extraction with Important blocks.

Next steps

Brunch adoption needs catalog admission and continuation tests against its real transport. User-guide screenshots remain pending for AI assistant › Experiments from chat and Experiments › Experiments created by the assistant.

Test coverage

• Core suites cover schema bounds, invalid requests, and correlated diagnostics
• Host suites cover captured inputs, ownership, completion, optimization refinement, and cancellation
• Chat suites cover tool-result identity, conversation changes, view navigation, card status, and accessible progress
• Website suites cover planning turns, streamed text before tool calls, returned result values, continuation, and cancellation during reply delays

Local verification passed: 995 React tests, 399 website tests, TypeScript, lint, package builds, formatting, architecture checks, and the 131-page docs build. Browser checks cover real simulation and optimization, distribution heatmaps and histograms, cancellation, protected controls, captured results, and normal/reduced-motion behavior.

How to test

Simulation

• Check out this branch and run turbo run dev --filter @apps/petrinaut-website
• Open /ai-experiments on the printed local URL
• Open AI assistant and ask How can I see the variation in this outbreak model?

  A streamed reply describes the baseline without starting an experiment.

• Send Run 256 simulations for 20 model-time units. and watch the blue card

  The card shows progress, then its metric and a completion notification.

• Select View experiment to inspect the metric heatmap
• Click a time step in the heatmap to inspect its histogram

Optimization

• Open /ai-experiments and select Edit > Settings > Simulation
  • Enable Parameter sweeps and In-browser optimization, then close Settings
• Open AI assistant and ask Can we compare different starting infection levels?
• Send Optimize for the lowest infected share.

  The purple card shows optimization steps and refinement. The final reply reports the selected starting share and returned metric.

• Select View experiment to inspect the best parameter value and distribution heatmap
• Click a time step to inspect its histogram across the completed runs

Ownership and cancellation

• Send optimize, then select View experiment while it runs

  Compute controls are read-only until the request finishes.

• Close the drawer and select Cancel on the active card

  The card becomes Cancelled and stops animating.

• Complete a new optimization, then explore another parameter value

  The earlier result in chat stays unchanged.

[vercel]

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

Project	Deployment	Actions	Updated
petrinaut	Ready	Preview	Sep 11, 2026 2:52am UTC
petrinaut-docs	Ready	Preview	Sep 11, 2026 2:52am UTC

2 Skipped Deployments

Project	Deployment	Actions	Updated
hash	Ignored	Preview	Sep 11, 2026 2:52am UTC
hashdotdesign-tokens	Ignored	Preview	Sep 11, 2026 2:52am UTC