I saw the Muse Spark release and wanted to reproduce the demo they showed (it looks cool). It's a simple example to show off agent canvas and automated coding agents:

Build an Agent Canvas version of a debugging demo that uses multimodal capabilities.

Hard model constraint: stay on the configured Sonnet profile. Do not switch to Opus or any other model/profile.

Demo requirements:
1. Build a small chat web app from scratch. Prefer plain HTML/CSS/JS plus a tiny built-in Node or Python server so the demo is robust without package downloads.
2. Include a local simulated assistant response path and an API-failure/fallback path. No paid/external API dependency.
3. Launch the app locally.
4. Use browser tools to open the app, interact with it, and take automated screenshots. Test one normal prompt and one failure/fallback path.
5. Inspect screenshots for user-visible failures or weak states. Describe what is visibly wrong and trace each issue to the relevant source code.
6. Patch the code.
7. Re-run and take follow-up screenshots to validate the fixes.