Turnback

io.github.MFaizR77Published by GitHub account MFaizR77v0.9.1Updated Oct 24, 2026

Undo a coding agent turn, including shell commands and untracked files.

VerifiedSTDIODesktop onlyFiles & StorageDeveloper Tools

Overview

AI-generated overview

Turnback snapshots a coding workspace before each agent turn and shell command, letting an assistant list, diff and restore its own turns.

What you can do
  • List recent turns with their prompts and inspect the steps, edits and shell commands inside a turn
  • Diff a turn to see what it changed, and compare or search across turns
  • Restore a whole turn or a point before a specific step, with a plan first and then the change
  • Bring back individual deleted files or folders and see which turn and prompt wrote each line
  • Record any command as an undoable turn and export history, including prompt attribution for a pull request
When to use it
Useful when an AI coding agent runs destructive or generative shell commands such as rm -rf, mv, code generators or git checkout, where built-in checkpoints only track edit tools. Also handy for keeping one undo history across several agents or for attributing lines and branch changes to the prompts that produced them.
Requirements
Runs locally over stdio as an npm package (turnback) with Node.js 22+ and Git 2.25+. Hooks and MCP are installed per agent via the CLI or a plugin marketplace; no API keys or environment variables are declared.
Before you install
Snapshots are stored in a shadow git repo under the home directory and cover tracked, untracked and gitignored files up to 5 MB, so secrets such as .env files may be copied there. Restore, undo and redo change files in the workspace, and export --commit writes to the project's .git. Old turns are cleaned up automatically, and very large workspaces fall back to edits-only mode where shell commands are unprotected.

Installation

In SourceWeft

  1. Open Turnback in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

[Turnback]

Undo for AI coding agents, even after rm -rf.

One undo history for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and Antigravity CLI.

[npm] [CI] [License: MIT]

[An agent runs rm -rf src .env; turnback undo brings both back]

A real Turnback run. The agent is simulated with Claude Code's hook payloads; the rm -rf and every Turnback line are real. Recreate it with python scripts/demo/make_demo.py.

Why

Agent checkpoints track the agent's own edit tools. Claude Code's documentation says it plainly: "Checkpointing does not track files modified by Bash commands." So when an agent runs rm -rf, mv, a code generator, or git checkout -- ., rewind cannot bring those files back.

Turnback snapshots the workspace before every shell command and edit, into a shadow git repo under ~/.turnback:

  • Undo a whole turn, or go back to just before one step of it.
  • Tracked, untracked, and gitignored files (like .env) up to 5 MB are covered.
  • Nothing writes to your project's .git unless you run turnback export --commit, and hooks never block the agent.
  • One history across six agents. Through MCP, the agent can list and diff its own turns and restore them in two steps: a plan first, then the change.

Install

In Claude Code:

bash
claude plugin marketplace add MFaizR77/turnbackclaude plugin install turnback@turnback

Then use /turnback:turns, /turnback:diff-turn, /turnback:undo, and /turnback:report.

In Codex (available in the next npm release):

bash
codex plugin marketplace add MFaizR77/turnbackcodex plugin add turnback@turnback

Start a new session, review and trust Turnback's hooks in /hooks, then use $turnback:turns, $turnback:diff-turn, $turnback:undo, and $turnback:report.

The plugin needs Turnback 0.9.1 or later. To try unreleased plugin changes from a checkout, see local testing.

For Codex, Cursor, Gemini CLI, OpenCode, Antigravity CLI, or the CLI on its own (Node.js 22+ and Git 2.25+):

bash
npm install -g turnbackturnback install all            # user-level hooks and MCP for every agentturnback install codex --project  # or one agent, in this project only

Claude Code, Codex, OpenCode, and Antigravity CLI have been tested live; Gemini CLI and Cursor are covered by tests built from their documented hook payloads. Details per agent: installation.

Compared with Claude Code's /rewind

/rewindTurnback
Edits made by the agent's edit tools✅✅
Files changed by shell commands (rm, mv, codegen)❌✅
Untracked and gitignored filesonly files its edit tools changed✅ up to 5 MB
Other agents❌Codex, Cursor, Gemini CLI, OpenCode, Antigravity CLI
Go back to a point inside a turn❌✅ --before-step
Rewind the conversation✅❌ files only

They work side by side: rewind the conversation with /rewind, and the files with Turnback.

Everyday use

bash
turnback --help                     # commands, options, and quick-start examplesturnback list                       # recent turns with their promptsturnback diff <turn>                # what a turn changedturnback undo --dry-run             # show the planturnback undo --yes                 # undo the latest turnturnback redo --yes                 # changed your mindturnback steps <turn>               # each edit and shell command in a turnturnback restore <turn> --before-step 3 --yesturnback mark "before migration"    # a checkpoint of your ownturnback run -- npm run codegen     # any command as an undoable turn, no agent neededturnback recover src/app.ts         # bring back one file (or a folder) an agent deleted turns agoturnback blame src/auth.ts          # which turn and prompt wrote each lineturnback report --pr --apply        # add which prompts wrote this branch to its pull requestturnback du                         # Turnback's disk use for every workspaceturnback ui                         # timeline in the browser

Without an agent

turnback run records any command as a turn, so a code generator, a codemod, or an agent without hooks gets the same undo. blame shows which turn wrote each line, and recover brings back a single file:

[turnback run records a code generator; blame shows its lines; recover brings back the file it deleted]

All commands, including log, search, report, compare, and export: commands. More: restore, MCP, scope and limits.

How it works

The hook before a mutating tool takes the turn baseline. The hook before a shell command captures the whole tree; later edits capture the affected paths. The end of the turn captures the result. The initial snapshot is warmed in the background on install and session start. Hooks always let the agent continue when recording fails; a failed, skipped, or unprotected state is visible through status.

Tracked, untracked, and gitignored files up to 5 MB are covered. Build output and dependency directories are excluded. Extra rules use gitignore syntax in .turnbackignore; global rules go in ~/.turnback/config.json as {"exclude":["pattern"]}.

Workspaces above 100k files or 2 GB switch to edits-only mode: only paths touched by edit tools are snapshotted, shell commands are recorded as unprotected, and restore touches only recorded paths. Turns older than 7 days and outside the last 50 turns are cleaned up automatically, at most once a day, in every workspace and not only the one in use. Data of a workspace that was deleted is removed after 30 idle days; set {"autoClean": false} in ~/.turnback/config.json to keep it.

Contributing

Issues labeled good first issue are small and say which files they touch, and support for more agents is welcome. Start with CONTRIBUTING.md and the architecture overview. Questions and ideas go to Discussions.

Contributors

[Contributors]

Every merged pull request is credited in the release notes.

Development

ModuleContents
src/cli/main.ts: CLI and hook entry point (built as dist/cli.js); args.ts: argument parsing
src/core/recorder.ts: snapshot rules per event; store.ts: journal, snapshots, turns, steps, marks, status, gc; restore.ts: restore planning and execution, undo, redo; format.ts: text output; quote.ts: shell-safe arguments; journal.ts, lock.ts, config.ts, types.ts
src/git/shadow.ts: shadow git repo wrapper
src/workspace/workspace.ts: workspace scanning and exclusion rules
src/agents/adapters.ts: each agent's hook payload → HookEvent; install.ts: hook and MCP installation per agent; opencode-plugin.ts: generated OpenCode plugin
src/mcp/server.ts: stdio MCP server
src/ui/server.ts: read-only local server for turnback ui; page.ts: its single page
bash
npm run checknpm run bench

The benchmark creates a temporary 10k-file repo and reports latency without making it a strict CI gate. CI runs the tests and benchmark on Windows, macOS, and Linux × Node 22/24.

From a clone, run npm ci && npm run build and use node dist/cli.js in place of turnback.

Releasing

Set the same version in package.json and in both version fields of server.json, commit, then push a v<version> tag. The Publish workflow checks that the versions match, runs the tests, publishes to npm through trusted publishing (provenance is attached automatically), and publishes server.json to the MCP registry as io.github.MFaizR77/turnback.

Also update src/core/config.ts, both plugin manifests, and the npm ranges in .claude-plugin/marketplace.json and .agents/plugins/marketplace.json.

npm trusted publishing is configured on the package page at npmjs.com (repository MFaizR77/turnback, workflow publish.yml), which needs the package to exist. The first version is therefore published once by hand with npm publish --access public; pushing its tag afterwards skips npm (the version exists) and only publishes to the MCP registry. Later tags do both.

License

MIT

Source: README.md at commit 2b468da

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

6
  1. v0.9.1LatestOct 14, 2026
  2. v0.9.0Oct 1, 2026
  3. v0.8.0Sep 30, 2026
  4. v0.5.0Sep 30, 2026
  5. v0.4.0Sep 30, 2026
  6. v0.1.0Sep 29, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts