Conversation handoffs
Know when to checkpoint a long conversation and carry the right context into a fresh one.
On this page
Use the checkpoint as a signal
Session details show a conversation checkpoint based on observed input-token size, request count, age and spend. It suggests continuing, preparing a handoff, or starting a fresh conversation. It is a heuristic, not a measurement of reasoning quality.
Leash does not retain your prompts or model responses. The handoff starts with metered session context; you supply the goal, completed work, decisions and next steps so the next agent has the information it needs.
Save a private HANDOFF.md
Choose a session, or omit its ID to use the latest one. The CLI creates a mode-0600 Markdown file and refuses to overwrite an existing file. Use --json to inspect the recommendation without writing a handoff.
leash handoff SESSION_ID \
--output ./HANDOFF.md \
--goal "Finish the payment failure path" \
--completed "Webhook signature checks pass" \
--decisions "Keep billing failures read-only" \
--next "Add the expired-trial browser test"
# Assessment only:
leash handoff --jsonStart the next conversation
- Review the handoff and fill in file paths, decisions, open questions and the exact next action. Remove credentials or sensitive text before sharing it.
- End the old agent run yourself. Creating a handoff does not kill or replace an active conversation.
- Start a fresh run with leash run and paste the continuation prompt printed by the CLI, pointing the agent at your handoff file.
leash run --cap 5 -- claude
# Or: leash run --cap 5 -- codex