Skip to main contentNew
- Token budgets now cap both input and output tokens in adaptive chats so turns stay within the model’s context window.
- Added a developer-only override that lets you test different context window presets without rebuilding the app.
Improvements
- Usage bar and overflow banners surface projected token totals alongside their limits, making pressure on the context window easier to track.
- Adaptive chat flow recalculates usage with the new caps to give more accurate guidance before a turn sends.
Fixes
- Stream error handling aligns with the updated token budgets, avoiding false failures when retries pick up a revised cap.