hi gabriel, the problem you are trying to solve is valid, but both claude-code and codex natively support local models ran through ollama. i've tried doing this before, and SLMs still fail to do basic tasks because they are put in a harness that has too many features and tools.
i've had really good success though in using them in minimal harnesses like pi-coding-agent. without an open-source repo it would be difficult to guess/assume what your approach is, but i hope you're thinking in the same direction and not trying to reinvent the wheel.
i've had really good success though in using them in minimal harnesses like pi-coding-agent. without an open-source repo it would be difficult to guess/assume what your approach is, but i hope you're thinking in the same direction and not trying to reinvent the wheel.
best, shaurya