Carry work across Claude Code, Codex and Cursor

Prepare the next host before you switch

Keep the project, .nightshift/ state, branch, and local commits available to both hosts. Write the active objective and Verify lines in punch-list.md; record decisions and exact next action. Stop the current shift deliberately, check its lease and open boxes, then open the same project in the receiving host and Start there. Do not start two hosts on one contract simultaneously.

• The approved list and checks are authority, not a chat summary.
• The parking lot carries owner decisions and reversible defaults.
• The snag log records failures; receipts and commits show what actually changed.

Documented path: Claude Code to Cursor

In the public AdaptTable example, Claude Code armed an eight-item contract on 12 August 2026. The owner issued a stop-work order after the Claude allowance ran out. Ten minutes after that stop, Cursor resumed the same branch, punch list, and definition of done. The contract reached its final tick on 14 August, and owner review and additional fixes preceded merge of PR #353.

• The public result records six roadmap phases, 67 commits, and 57 closed issues after owner review.
• 45 hours 7 minutes is elapsed contract continuity, not uninterrupted compute time.
• The example proves this handoff; Codex is a supported host, but this case does not prove a Claude Code-to-Codex result.

Review before another handoff

Check the current branch and working tree, the item receipts, tests, parking lot, and open boxes. A recovered process can own a lease even when an old panel looks stale; use Status or Doctor and follow the host recovery guidance before interacting. The next host inherits durable state, not a guarantee of identical transcript history or permissions.