Changelog: 2026-03-15
Chat gained safer, more controllable agent runs with in-flight tool approvals, improved budgeting across sub-agents, and clearer streamed status—plus major upgrades to search and retrieval quality.
- Added in-chat tool approvals so certain actions can be approved or rejected while a response is still running.
- Improved agent reliability with better budgeting across sub-agents and more accurate handling of failures during streaming.
- Introduced optional streamed “thinking” that can be shown or hidden with a user preference.
- Upgraded header search with faster keyboard access, clearer empty states, and richer result context.
- Improved retrieval quality with smarter post-processing (reranking, deduping, similarity filtering, and token budgeting).
- Modernized the scraping and analysis pipeline foundation for more consistent runs and easier reuse across components.
- Chat controls: When an assistant needs to use a gated tool, execution can pause and request approval, then continue based on approve/reject feedback.
- Agent budgeting & correctness: Usage limits now account for nested agent work, and tool/agent states are more consistently marked when errors or disconnects occur mid-stream.
- Reasoning visibility: “Thinking” can be streamed to the UI and displayed in a collapsible view, with a persistent per-user toggle.
- Navigation & search: Pressing
/focuses search, the dropdown provides clearer calls-to-action when results are empty, and result cards include more helpful metadata. - Retrieval improvements: Added a post-retrieval pipeline to reduce duplicates, prioritize better matches, and stay within token limits for more focused answers.
- Scraping & analysis pipeline: Core scraping and analysis components were refactored into a more unified architecture with improved run tracking and consistency.
- General reliability: Build and release stability improvements reduce avoidable install/build failures and help ensure changes deploy more predictably.