Repository navigation
v1.5.2
Release Notes - v1.5.2
Features
Tool Result Persistence
The assistant can now remember tool results across conversation turns, eliminating redundant data fetching. When you ask follow-up questions, the model will reuse previously retrieved information (such as emails or documents) instead of fetching them again.
Key details:
- Tool call history is automatically stored in conversation metadata
- Recent tool calls are replayed with full context within token budget limits
- Older tool calls use compact summaries with cached results
- Opt-in feature: Enable in Settings > LLM > Tool Context Replay (disabled by default)
- Works retroactively for existing conversations once enabled
Error Reporting Integration
Added Sentry integration for improved error tracking and monitoring. This enhancement helps the development team identify and resolve issues more quickly.
- Configured via
logging.sentry_dsnsetting orSENTRY_DSNenvironment variable - Automatically initialized at application startup
Improvements
- Enhanced conversation context management with configurable token budgets
- Improved memory efficiency for long-running conversations
Dependencies
- Updated proxy-addr dependency to address security updates