FE-1668: Demonstrate AI experiment and optimization integration

[kube]

Important

Experimental optimization
Optimization requires Parameter sweeps and In-browser optimization. Simulation is available without these flags.

Summary

Add an /ai-experiments demo with scripted AI replies and real browser computation. The conversation proposes an experiment, starts it on request, and reports the returned results.

The example uses a saved SIR model with distributions. The integration guide separates Petrinaut's execution and chat responsibilities from Brunch's tool selection, transport, and agent behavior.

Experiment

experiment.mp4

Optimization

optimization.mp4

Links

• Ticket FE-1668 (internal)
• Ticket FE-1484 (internal)
• Docs AI client integration
• Docs AI-created experiments › Try the demo

Blocked by

• #9678

Changes

• Stream planning replies before accepting a command to run or optimize
• Derive completion messages from actual metric values, selected parameters, and run counts
• Demonstrate heatmaps and histograms for simulation and optimization
• Document the steps for Brunch adoption and its separate production integration work

Known issues

The SIR example is illustrative and uncalibrated. Searching initial infection levels demonstrates parameter search. Brunch adoption requires its own catalog registration and transport test, tracked in FE-1484.

Test coverage

The stack includes a fix for the HASH Edge deployment: AI schemas no longer import the worker backend. A bundler regression covers the dependency boundary, and all 52 integrated validation tasks passed again after the fix.

After the 15 September restack onto main a898a13810, core, editor, website, and Brunch builds, unit tests, TypeScript, and lint passed on the integrated stack (52 tasks). Architecture validation, the documentation site build, and changed-file formatting passed.

• Demo transport tests cover planning, explicit launch commands, streamed replies, cancellation, and returned results
• The recordings show live progress, completed results, and metric distributions

Browser verification completed 256 simulation runs, opened the heatmap and a time-slice histogram, and confirmed the fixed title is read-only. No console errors were reported. Optimization and cancellation are covered by tests; browser optimization was not rerun at the user's request. Existing recordings predate this restack.

How to test

• Run turbo run dev --filter @apps/petrinaut-website
• Open http://localhost:5173/ai-experiments, using the printed port if it differs
• Open the AI assistant and ask How can I see the variation in this outbreak model?
• Send run after reading the proposal
• Select View experiment, then click a heatmap time step to inspect its histogram
• Open Settings and enable Parameter sweeps and In-browser optimization
• Ask to compare starting infection levels, then send optimize
• Inspect the running card and the best point's distribution after completion
• Start another optimization and select Cancel

  The card stops updating and shows cancellation.

[vercel]

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

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 15, 2026 12:31pm UTC
petrinaut	Ready	Preview	Sep 15, 2026 12:31pm UTC
petrinaut-docs	Error		Sep 15, 2026 12:31pm UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 15, 2026 12:31pm UTC

[lunelson]

Approved by Lu Nelson following an agent-assisted review of this PR against its parent and FE-1668. No actionable introduced defects found in the scripted transport, explicit launch commands, cancellation, or result-derived replies. The distinction between this demo and production Brunch integration is clear.

Verification on the integrated four-PR stack: core package build passed; 231 targeted tests across 11 files passed, including all 26 demo transport tests. Live-browser simulation and optimization were not rerun during this review; the recordings predate the rebase, as noted in the PR.