rewindagents

Undo for AI coding agents. rewindagents records every tool call your coding agent makes and keeps a copy of each file before a step changes it, including files only a shell command touched (rm -rf, sed -i, a script). When a run goes wrong you see what it changed and take back all of it, one step, or one file. It runs on your machine only: no account, no network, nothing but the Python standard library.

Quick start

The flight recorder for coding agents: see exactly what your agent did, and take back any part of it.

pipx install git+https://github.com/Johnbjoseph-cybersec/rewindagents && cd your-project && rewindagents init

init looks for the agents it supports (Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, Droid, Qwen Code): their config directory in the project or your home directory, or their command on PATH. It prints the plan, asks [y/N], and adds the hook to each one it found, one line per agent: installed, already there, replaced or not found. It never adds a second copy and keeps every hook already in the file. --yes skips the question. --global writes each agent's user config instead, so every project is recorded. Without --yes and without a terminal to ask on, it only prints the plan. --agent claude sets up only that agent (repeat it for more). In a git repository the config files init creates are listed in .git/info/exclude, this clone's own ignore list, because they name this machine's Python and would not run on a teammate's; a file git already tracks is changed and named instead. rewindagents init --undo takes it all out again and keeps what was recorded.

Then use your agent as usual and:

rewindagents summary                   # what it changed: each file, lines added and removed, which steps
rewindagents replay                    # what it did, one line per step
rewindagents diff 12                   # what taking it back to step 12 would change
rewindagents rewind 12 --only src/db.py   # take back just that file

rewindagents status says in plain words whether recording works right now: it runs each hook once in a scratch folder and checks that it wrote a record.

After a bad run

Three commands, in the order you need them. summary prints the last two filled in for the session.

rewindagents summary          # which files are different now, by how much, and which steps did it
rewindagents revert 7         # take back step 7 alone, keep every later step and hand edit
rewindagents rewind 0         # put every file back as it was before the session's first step

rewind and revert refuse, change nothing and exit 1 when they would overwrite a change made after the agent's step (your own edit, for example); the message names the file and the command that shows what would be lost. Every rewind is itself a step, so it can be undone, and it prints the command that undoes it.

Here is a real session: an agent's last two steps delete the tests and overwrite the config, replay lists the fourteen steps, rewind 12 puts the files back.

brand/demo/build.py runs those fourteen agent steps through the hook in a temporary project, then the four commands shown, and draws the frames from what they printed; tests/test_demo.py plays the same session again and fails if any line differs.

What it cannot put back

Files under a directory git ignores as a whole (node_modules/, build/), symlinks, files over 20 MB, changes outside the project, and the shell steps of a project with more than 20,000 files (status warns about that one, and replay marks such a step). Details under Claude Code hook, Codex and Cursor hooks and Replay, rewind, branch.

Claude Code's own checkpointing docs (read 2026-11-12) name the gap this fills: "A file that only Bash commands changed isn't tracked and stays as it is." (its /rewind does undo Bash changes to a file Claude had already edited), and, of subagents, "Claude Code usually doesn't capture those edits in your session's checkpoints." rewindagents copies the project's files before every shell command and scans them after, so a sed -i or an rm is one more step you can rewind, and Claude Code runs the same hooks for a subagent's tool calls (hooks docs). A live session in which a subagent made eight steps, three of them edits, recorded and checkpointed all of them (ux-review/DOGFOOD.md). Every tool call goes into a tamper-evident, hash-chained log you can verify later.

Status: v0.0.8, pre-release. Built and tested today: the recorder and its CLI, hooks for seven coding agents (Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, Droid, Qwen Code) that record every call and checkpoint the files a step changes, summary, replay, rewind, revert and branch, fork and compare, export (a session as one HTML page), sessions, bundle and import, gc, a local timeline UI, and a pip-installable package with a rewindagents command. Not on PyPI yet. The Claude Code hook has run in real sessions; the Codex and Cursor hooks are built from the vendors' hook documentation and tested with the documented payloads, not yet in a live session. macOS and Linux; Windows is not supported yet (see Recorder).

Rewind to any step and let another agent take it from there: Fork a step, compare the runs.

Install

Python 3.10 or newer (declared in pyproject.toml; tested on 3.15, macOS; the full suite also passed there on 3.10, 3.11, 3.12, 3.13 and 3.14 on 2026-11-14). No dependencies beyond the standard library. From the repository:

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗