Changes
- Improve agent context handling with microcompaction plus full LLM-based compaction to preserve progress on long research sessions.
- Add concurrent execution for safe read-only tools to reduce waiting time during multi-tool runs.
- Add mid-run message queueing so new user messages are picked up while the agent is still working.
- Improve model invocation with message-array streaming and fallback behavior for providers without stable streaming.
- Update tool registry metadata with compact prompt descriptions and per-tool concurrency safety.
- Improve CLI/chat UX with queued-message indicators, compaction events, and cleaner tool spinner lifecycle.
- Update core dependencies (LangChain stack, Playwright, and related packages) for newer provider/runtime behavior.
Fixes
- Fix model/provider sync after runtime model switch so subsequent turns use the selected config.
- Fix cron tool schema examples to avoid JSON-curly-brace parsing issues in tool descriptions.
- Fix early exit behavior when interrupted (Escape) and prevent stale tool spinners in interrupted runs.