Why OASIS

A coding agent is a model plus a harness. The model decides what to do; the harness decides what actually happens to your repository. Most of what people hate about coding agents is harness failure: tool calls that don’t parse, edits that miss by one space, rm -rf in auto mode, “done” without running a test, a function replaced by // ... rest unchanged, a terminal that looks frozen for two minutes.

OASIS fixes each of those in the runtime, so the fix holds whichever model you connect.

Any model, really. Claude, GPT, Gemini, DeepSeek, OpenRouter, Bedrock, Vertex, Azure Foundry, or a local Ollama, LM Studio, llama.cpp or vLLM server. Small models get a compact prompt, a 13-tool profile and tool-call repair, so even a 4B local model drives the tools; whether its code is right is still up to the model. In our live runs on a 2-CPU cloud machine with no GPU, Qwen3-4B passed two of four hidden-test tasks (parse-boolean-flag and dedupe-names). clamp-integer failed in all three tries. safe-task-id passed once in two tries, but that code fails the stricter verifier added afterwards ([field…