Every fresh coding-agent session can turn the user into the synchronization layer. You have to restate what was completed, what is still open, which decision changed, and what the next agent must not redo.
We built a small Apache-2.0 experiment in Lians around that exact problem. It runs locally through MCP and SQLite, asks for no Cursor, Claude, or Codex API key, and returns a bounded current-state handoff instead of replaying the whole conversation.
The committed Claude-to-Codex fixture currently recovers 10/10 expected facts, presents 0 stale facts as current, excludes one superseded fact, and produces an estimated 231-token handoff. That is a bounded fixture result, not a claim that arbitrary live sessions extract perfectly.
Try the smallest useful Cursor test
- Install Lians through the Cursor MCP integration.
- In one chat, say:
Remember that this project uses Python 3.12 and pytest. - Open a fresh chat in the same project.
- Ask:
What Python version and test runner does this project use?
- Repository: GitHub - Lians-ai/Lians: Evidence-backed proof of done for Claude Code, Codex, Cursor, and other AI coding agents. Run real checks, bind results to current Git state, and know what is ready for human review. · GitHub
- Cursor setup: Lians/integrations/cursor at master · Lians-ai/Lians · GitHub
The cross-agent reproduction is under experiments/cross-agent-continuity in the repository.
The free local path needs no Lians account, Docker service, desktop application, or provider API key. The automatic Claude SessionEnd to fresh Codex flow remains a source beta; the published package supports the smaller explicit remember, recall, inspect, correct, and forget loop.
I would value failure reports more than compliments, especially around project scoping, stale facts, redundant work, and which details a fresh Cursor session actually needs.