Open Questions and Risks

Questions

• Which model/browser/device combinations are officially supported?
• What are the measured first-load and warm-cache latencies?
• How often do large model downloads fail or exceed available memory?
• Are model CDN assets pinned, integrity-checked, or mirrored?
• Should conversation export or persistence be added?
• Which tools are safe enough to enable by default?

Risks

Risk	Impact	Mitigation
WebGPU or memory failure	User cannot load a model	Smaller presets, capability checks, remote fallback
Worker/audio race	Lost transcript or stuck status	Integration tests and explicit lifecycle state
Dependency/model compromise	Browser privacy boundary weakened	CSP, auditing, pinning, trusted mirrors
Cache eviction	Repeated multi-GB downloads	Cache diagnostics and user-visible storage guidance
Tool prompt injection	Incorrect local action	Strict schemas, allowlists, confirmations

Practical next steps

1. Add Playwright coverage using mocked STT/LLM/TTS boundaries.
2. Add a model-loading compatibility matrix to the README.
3. Reduce production console payloads and add privacy-safe diagnostics.
4. Document cache clearing and recovery after failed model loads.

Verification checklist

• Test every default preset on supported Chrome/Edge desktop hardware.
• Test iOS fallback with STT and TTS disabled/enabled combinations.
• Verify no old worker/model emits after reset or model switch.
• Measure first-token and first-audio latency for each engine family.
• Verify CDN failures produce actionable retry/reset UI.
• Add privacy-safe CSP and dependency/model integrity policy.
• Decide whether local conversation export is a product requirement.
• Confirm experimental tool behavior before enabling it by default.