Handoff

The train problem, properly: you're an hour into a local Claude Code session, it's mid-refactor, and you have to leave. Starting a fresh cluster session loses the conversation. So don't:

# stop the local claude first (ctrl+c) — two copies of one transcript
# tell two different stories
tiny handoff

Run from the project directory. It ships three things into a new session:

1. The working tree, exactly as it sits — uncommitted changes and .git included. No push, no stash, no credentials involved: the tree streams over the Kubernetes exec API into /workspace/repo.
2. The transcript — the local Claude Code session file for this directory, placed where the cluster session's claude --continue will find it.
3. A nudge. The session was created waiting; once the files land, its agent starts on the resume path, reopens your conversation, and keeps working. In our test its first act was git status && git diff to orient itself, which is exactly what you'd do.

Then close the laptop. tiny shows it working; tiny attach <name> drops you back into the same conversation from any machine.

Honest notes

• Stop the local claude before handing off. The command warns if the transcript was written seconds ago, but it can't see your terminal.
• Claude Code only. The command reads the local Claude Code transcript, so a Codex session cannot be moved this way. Everything after the move works for either agent.
• The tree ships whole, with no .gitignore filter and no exclude list. A 10GB repo ships 10GB, and node_modules, build output and any .env travel with it — tidy the directory before handing off a monorepo with models in it.
• The return leg is tiny pull: it brings the workspace back to a new local directory, and refuses to write into one that already has files in it.
• No local conversation to carry? tiny new --dir . ships just the folder — same transfer, no transcript.