Just wanted to share something cool I've been working on for the past half a year.
CodeVROOM[0] is a symbol-level AI editor designed for very fast incremental edits to large projects.
Instead of including entire files with your editing request, it trims them to include just the relevant parts, while preserving the overall structure, so the models won't get confused. It can also automatically ask the model what other symbols are relevant to a particular edit, and restart the conversation with these symbols. Here's an example [1] how a couple of context discovery steps quickly get from a hallucination-ridden answer to a straight-to-the-point implementation. And because the model works with heavily trimmed source files, it works very fast (1-2 seconds per step with Cerebras platform) and has negligible token costs (typically, under a cent per edit).
The main difference from CLI tools is to give the user full control over what's going on. You can step the editing session back, see what symbols the model deemed relevant, edit them, and retry. Or you can retry individual steps, switch between models at any point, or do special steps like "discover more context" or "expand my original instructions, giving several options". Most steps only take a few seconds with smaller models (LLaMA on Cerebras) or about a minute on larger models (Claude), so it's very easy to review what the model is doing, and steer it into the correct direction. You can also do things like ask Claude to expand your instructions, and then let LLaMA implement them.
CodeVROOM is built from scratch using .Net (WPF on Windows, Avalonia on Linux/Mac), so it has some unique features that other editors don't have. E.g. the change reviewing logic [2] is integrated with the outline, so you always have an overview of the added/removed/edited members, and can approve/reject edits in bulk. Or you can collapse series of unchanged members to see all edits at a glance (but with full context unlike plain diffs), or switch between inline and side-by-side modes.
The main use case is routine edits and refactorings that are not handled by the existing refactoring tools, and take several minutes to do by hand. Even small AI models handle these very well, and getting reviewable results from a consise prompt in a second or two is just way less distracting than doing it by hand. There are some examples of real-world edits here: [2].
There's also support for AI techniques where you provide plain-text instructions how to do a particular recurring task, and can then reference it from a prompt (port size-related properties per @PropertyPorting) and edit templates where you can write instructions for common edits (e.g. add null checks to a function) and then invoke them by clicking on a link above the function.
CodeVROOM supports 10 common cloud providers and local models via Ollama. It has old-school perpetual licenses with 1 year of free updates, there is no telemetry, and no lock-in. The product is the GUI, not your data.
It is stil early in development with tons of more features coming, but it can already save you some time side-by-side with your main IDE. The trial will remain unlimited until more IDE features are ready.
[1] https://sysprogs.com/CodeVROOM/documentation/examples/connec...
[2] https://sysprogs.com/CodeVROOM/documentation/features/diffs/
No comments yet.
