VB Harness is a Ruby tool for getting local language models to make code changes while limiting common failures such as invented APIs, faulty edits, malformed responses, and small context windows. Its central idea is that the model should not judge its own work: the project’s type checks, lint, build, and tests determine whether a change is accepted. A request is split into small, file-limited tasks, and the plan is checked before implementation. Each task is applied as exact edits, tested, and sent back for fixes if checks fail. A final reviewer can report a bug only when it can demonstrate the problem with a failing test.
Work takes place in a separate Git worktree and branch, leaving the user’s checkout untouched; the harness does not push changes or install dependencies without permission. It supports OpenAI-compatible model servers such as LM Studio, Ollama, llama.cpp, and vLLM, with a browser app and command-line interface. Model calls and their costs are tracked by role. AstroJS is the first supported stack, and the tool has no Ruby gem dependencies, but it is explicitly in early development. A showcased run built a four-part blog feature on a laptop with a 6 GB GPU, taking three hours and 28 minutes - an illustration of what the workflow can accomplish, not a guarantee of speed or correctness.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.