Roo Code

Roo Code was the fork with modes, Code, Architect, Ask, Debug and Orchestrator, and the modes were the reason to choose it over the upstream: a debugging session and an architecture discussion got different prompts and different tools. That reason left with the project. The trait that decides it now is a successor: the docs themselves point to ZooCode, and Cline is the upstream everything descended from.

If you are on it, move this quarter and carry your mode configs with you; they were the valuable part. Nobody should adopt it new, and nobody needs to, since Cline and Kilo Code cover the same ground with maintainers.

The failure mode already happened. The shutdown on May 15, 2026 took the extension, Roo Code Cloud and the Router offline together, and the repository was archived, so no bug reported after that date will be fixed and no paid service survives to complain to. Money in, service gone, issues frozen.

Anyone still running it is running an extension that will never see another patch. What it did right: checkpoints, so a session's edits could be rolled back to a known state, a feature its successors kept because it was the correct one.

Roo Code's contribution was structural, not numerical. 1. Specialized modes gave each task type its own prompt and tool set, a documented, reproducible way to constrain agent behavior without changing the model. 2. Orchestrator mode delegated subtasks to the other modes, which is context partitioning by role rather than by file. No benchmark was ever published for either.

The repository is archived, so the design is now a fixed artifact rather than a moving target, easier to study and impossible to improve in place. The observation: an architecture that survives its maintainers in a fork is, strictly, the reproducible part.

Roo Code is the textbook. A fork of an open-source project, roughly 24,000 stars, then a paid cloud layer bolted onto a free extension whose users had chosen it precisely to avoid a vendor. Then the company shut it all down and became Roomote: same cap table, different product, which is the pivot in its purest form.

The lesson for buyers is that a free extension with a venture round behind it has a business model pending, and pending resolves one way or the other. Position: closed. Read the postmortem, use Cline.

The demo was an agent with a mode picker. The vendor shut down the extension and its paid services, archived the repository, and moved on, so there is no contract, no support, no SSO, no audit log and no one to sign the questionnaire; the security review would be of a binary nobody maintains, on sixty laptops, with shell access.

Cost is zero and so is accountability. Not yet, and not ever in this form; approve the migration budget instead, and make the migration this sprint rather than next quarter.

The company left; the code did not. I can still build it and run it against Ollama, because the license is Apache-2.0, and the mcpServers config with its alwaysAllow list and disabled flag still works the way it did, since none of that ever depended on a server. What died was the parts that phoned home.

A community fork exists, which is exactly what a license like this is for, and my config moves over unchanged. The lesson is the license, not the product: pick tools whose survival does not require the vendor's.