Tumble Code

Your AI-Powered Dev Team, Right in Your Editor

Tumble Code is a community-maintained fork of Roo Code. The original Roo Code project is no longer actively maintained by its authors; Tumble Code carries the codebase forward under independent stewardship. See the Lineage section below for full attribution.

What's new in this release:

• Add GPT-5.5 support via the OpenAI Codex provider.
• Add Claude Opus 4.7 support on Vertex AI.
• Add previous checkpoint navigation controls in chat so you can step back through prior checkpoints more easily.

What Can Tumble Code Do For YOU?

• Generate Code from natural language descriptions and specs
• Adapt with Modes: Code, Architect, Ask, Debug, Vision, and Custom Modes
• Refactor & Debug existing code
• Write & Update documentation
• Answer Questions about your codebase
• Automate repetitive tasks
• Utilize MCP Servers
• Run with local inference engines (Ollama, LM Studio, llama.cpp) so your code never leaves your machine and you pay nothing per token

Modes

Tumble Code adapts to how you work:

• Code Mode: everyday coding, edits, and file ops
• Architect Mode: plan systems, specs, and migrations
• Ask Mode: fast answers, explanations, and docs
• Debug Mode: trace issues, add logs, isolate root causes
• Vision Mode: describe screenshots, diagrams and charts for models that cannot see images (pin a vision-capable model to it)
• Custom Modes: build specialized modes for your team or workflow

Resources

• Documentation: Tumble Code documentation is a work in progress; the architecture notes for contributors live in docs/.

Local Setup & Development

Requirements: the extension runs in VS Code 1.102 or newer. Development needs Node.js 22 (the exact version is in .nvmrc) and pnpm 12 (the exact version is packageManager in package.json; an installed pnpm 10 or newer switches to it on its own, or run corepack enable).

1. Clone the repo:

git clone https://github.com/krzychdre/tumble-code.git

2. Install dependencies:

pnpm install

3. Run the extension:

There are several ways to run the Tumble Code extension:

Development Mode (F5)

For active development, use VSCode's built-in debugging:

Press F5 (or go to Run → Start Debugging) in VSCode. This will open a new VSCode window with the Tumble Code extension running.

• Changes to the webview will appear immediately.
• Changes to the core extension will also hot reload automatically.

Automated VSIX Installation

To build and install the extension as a VSIX package directly into VSCode:

pnpm install:vsix [-y] [--editor=<command>]

This command will:

• Ask which editor command to use (code/cursor/code-insiders) - defaults to 'code'
• Uninstall any existing version of the extension.
• Build the latest VSIX package.
• Install the newly built VSIX.
• Prompt you to restart VS Code for changes to take effect.

Options:

• -y: Skip all confirmation prompts and use defaults
• --editor=<command>: Specify the editor command (e.g., --editor=cursor or --editor=code-insiders)

Manual VSIX Installation

If you prefer to install the VSIX package manually:

1. First, build the VSIX package:

   pnpm vsix
   

2. A .vsix file will be generated in the bin/ directory (e.g., bin/tumble-code-<version>.vsix).
3. Install it manually using the VSCode CLI:

   code --install-extension bin/tumble-code-<version>.vsix
   

Working over Remote SSH

When VSCode is connected to a remote machine through Remote SSH, the Tumble Code extension host runs on the server, not on your laptop. The extension runs Node code, so it defaults to the workspace extension kind, which means everything it does happens where the workspace lives.

Where state lives

• All persistent state (global storage: task history, API provider profiles, memory, artifacts) lives in ~/.vscode-server/data/User/globalStorage/ on the server, separately for each host you connect to.
• Task history and provider profiles are per-host. A task you finished on one server does not exist on another, and profiles saved on one host are not visible from another.
• The only state that physically lives on your laptop is secret storage (API keys): VSCode proxies it through the SSH channel to the client machine's keychain.
• The webview UI renders locally, but all extension logic and file access happen on the server.
• The extension must be installed "in SSH": in the Extensions view, use the button "Install in SSH: [host]" instead of installing it only in the local window. A purely local install does not run on the server.

Machine-scoped settings

The following settings are scoped machine. A value set in User settings on your laptop does not apply in a remote window; set them in the "Remote [SSH: ...]" settings tab (Settings editor, tab named after your SSH host) so they apply on the server:

• tumble-code.customStoragePath (a local disk path, resolved on the server)
• tumble-code.autoImportSettingsPath (same)
• tumble-code.debugProxy.serverUrl (a local proxy address)
• tumble-code.cloudApiUrl
• tumble-code.clerkBaseUrl

Remember that in a remote window localhost means the server, not your laptop, so addresses like http://127.0.0.1:... must point at something reachable from the server.

Settings that are not host-specific (for example allowedCommands, deniedCommands, commandExecutionTimeout) keep their normal scope and still apply from User settings or Settings Sync.

Diagnostic checklist

If task history or settings misbehave in an SSH window, check the server side first, because that is where the storage lives:

1. Disk space and quota on the server: df -h ~ and quota -s (a full disk or exceeded quota makes every write fail with ENOSPC or EDQUOT).
2. Filesystem type: stat -f ~/.vscode-server. Network filesystems (NFS) add latency that can make file locks time out.
3. inotify limits: cat /proc/sys/fs/inotify/max_user_watches and /proc/sys/fs/inotify/max_user_instances. vscode-server itself consumes thousands of watches, which can exhaust the limits and degrade file watching.
4. Extension logs: in the Output panel, check the Tumble Code channel and "Log (Remote Extension Host)". Look for storage errors (ENOSPC, EDQUOT, EROFS, EACCES), failed TaskHistoryStore initialization, and lock errors (Failed to acquire lock for, Lock compromised).
5. Effective settings: in the remote window, open Settings and compare the "Remote" tab against the "User" tab to see which value of a machine-scoped setting actually applies.

We use changesets for versioning and publishing. Check our CHANGELOG.md for release notes.

Disclaimer

Please note that the Tumble Code contributors do not make any representations or warranties regarding any code, models, or other tools provided or made available in connection with Tumble Code, any associated third-party tools, or any resulting outputs. You assume all risks associated with the use of any such tools or outputs; such tools are provided on an "AS IS" and "AS AVAILABLE" basis. Such risks may include, without limitation, intellectual property infringement, cyber vulnerabilities or attacks, bias, inaccuracies, errors, defects, viruses, downtime, property loss or damage, and/or personal injury. You are solely responsible for your use of any such tools or outputs (including, without limitation, the legality, appropriateness, and results thereof).

Contributing

We love community contributions! Get started by reading our CONTRIBUTING.md.

New to the code? Read the one-page architecture map before you change anything that crosses a folder boundary.

Lineage

Tumble Code is a community-maintained fork of Roo Code, originally authored by Roo Veterinary Inc. and published on the VS Code Marketplace as RooVeterinaryInc.roo-cline. In early 2026 the upstream team announced that they were stepping away from Roo Code to focus on their new product, Roomote. With the original maintainers no longer actively developing the extension, this fork was started to keep the codebase alive and shift the focus toward local inference engines — running with Ollama, LM Studio, llama.cpp, and other locally-hosted LLMs as first-class providers so users can use AI coding assistance without paying per-token or sending their code to a cloud provider.

What we kept from Roo Code:

• The entire codebase, Apache 2.0 licensed (see LICENSE for the full notice including the original Roo Code copyright)
• The agent architecture, mode system, MCP integration, provider abstractions, and webview UI
• Internal identifiers (roo-cline.* command IDs, view IDs, file names) so the rebrand stays additive rather than disruptive
• The original release history in CHANGELOG.md, preserved verbatim with a header note distinguishing the upstream and fork eras
• The translation work of the Roo Code community across 18 locales

What's different in Tumble Code:

• New marketplace identity: QUB-IT.tumble-code — install path is independent from the upstream extension
• The agent persona in chat copy is renamed from "Roo" to "Tumble"; the logo is a tumbleweed instead of a kangaroo
• Cloud features (auth, share links, telemetry, the LLM proxy) point at a self-hosted backend (see self-hosted-cloudapi/ in this repo) instead of the upstream project's servers. Default URLs are configurable via env vars
• Independent direction with a focus on local inference: Ollama, LM Studio, llama.cpp, and similar locally-hosted LLMs are treated as first-class providers, not afterthoughts. Cloud providers (Anthropic, OpenAI, OpenRouter, etc.) still work; they're just no longer the default expectation
• A first-launch settings migration bridge offers to copy your existing Roo Code configuration so you don't lose work

Credit:

The Roo Code project was the work of many contributors over multiple years; their commits remain in this repository's git history. If you contributed to upstream Roo Code, your work is still here. The fork carries forward your effort under a new name, not a clean-room reimplementation.

For questions about lineage, licensing, or attribution, please reach out through whichever contact channel Tumble Code publishes once it has one set up.

License

Apache 2.0 © 2025 Roo Code, Inc. — community fork © 2026 Tumble Code contributors

Enjoy Tumble Code!