Double-tap Esc to rewind the session — code, conversation, or both
Every prompt you send is a checkpoint. The rewind menu undoes a bad diff without touching the conversation, or rewinds the thread without touching the code.
Claude Code checkpoints your session automatically. Press Esc twice on an empty prompt — or run /rewind — and a menu lists every prompt you've sent this session. Pick one, then choose what to roll back: the code, the conversation, or both. Most people learn Esc as “interrupt Claude mid-turn” and stop there; the double-tap is the half of the feature that turns mistakes from cleanup jobs into a two-keystroke undo.
The mechanism is simple. Every user prompt creates a checkpoint, and Claude snapshots a file before each change its editing tools make. Checkpoints persist across sessions — close the terminal, resume tomorrow, the rewind menu still works — and get cleaned up with sessions after 30 days. The practical effect is on how you delegate: you can hand Claude a large, ambitious refactor knowing the floor under you is one menu away, instead of pre-committing to git before every risky ask.
Restore code is the one you'll reach for most. The edit went sideways but the conversation is still good: Claude has read half your codebase, understands the task, and you don't want to pay for that discovery again. Restoring code only throws away the bad diff and keeps everything Claude learned. No git stash, no checkout, no commit you made purely as a savepoint.
Restore conversation is the inverse: the files are fine, but the thread went down a wrong path — a misread requirement, a long tangent, an instruction you phrased badly and Claude has been faithfully over-applying since. Rewinding the conversation while keeping current code resets Claude's working memory to the point before the wrong turn. Usefully, the prompt from the message you selected is placed back in your input box, so the normal move is: rewind, edit the wording that caused the problem, resend.
Summarize from here is the sleeper. It's a targeted /compact: messages before the point you pick stay in full detail, everything from that point forward is replaced by a compact AI-written summary, and no files are touched. A verbose debugging session that finally found the fix is the classic case — keep your original instructions verbatim, compress the forty turns of log-reading into a paragraph, and reclaim the context window. The original messages stay in the session transcript, so the detail isn't gone, just out of context. You can type instructions to steer what the summary keeps.
The trap: checkpointing only tracks files changed by Claude's editing tools. If Claude runs rm, mv, or cp through bash, those changes are not snapshotted and will not come back on rewind — same for edits you make yourself outside the session. And it isn't version control: think of checkpoints as local undo and git as permanent history, not one replacing the other. One smaller gotcha — Esc Esc with text in the input clears your draft instead (it's saved to history; Up recalls it). The rewind menu only opens from an empty prompt.
The try-it below stages a deliberately disposable edit and reverses it in about ten seconds, which is all it takes for the double-tap to become reflex.
Make a scratch file, let Claude edit it, then rewind the code without losing the conversation.
mkdir -p /tmp/rewind-demo && cd /tmp/rewind-demo
echo 'def greet(name): return f"Hello, {name}"' > greet.py
claudeIn the session, ask:
Rewrite greet.py with type hints, a docstring, and logging.When the edit lands, press Esc Esc on the empty prompt. Select your prompt in the list, choose Restore code, then run !cat greet.py — the file is back to one line, but the conversation still remembers everything. On a longer real session, open the same menu and try Summarize from here on a midpoint message to see targeted compaction.