Marmel 0.9.0

Hi, Some weeks ago I announced marmel 0.1.0, an autonomous coding agent. Over the past two weeks huge efforts been put into stabilizing and improving the performance and reliability for in particular local models. I've spent hours tweaking and tuning the architecture to run decently with even small models such as gemma 4 12b, which despite all it's inherent problems so far managed to complete very task I assigned it. So therefor I now wish more people would like to battletest it, I already know it works great with cloud models such as deepseek v4/v4.1 flash, but would prefer more feedback from people using local models. To set expectations right, this is intentionally designed to be autonomous, to build prototypes (with hopefully good quality) I've put it through a lot of testing and my local rig been running pretty much 24/7 (as well as cloud testing, which I of course prefer as it's it's a completely different level of interactivity and responsiveness) And finally, I originally intended to save this "announcement" for the v1.0.0 release, but my ambitions are still only at a planning stage, and I think it's basically criminal considering how good my experience been with this for the past week not to let others test it. I think the read me of the project tells more than I can possibly "pitch" it here, of course the main goal is to make this as usable as possible so hopefully this post either can result in feedback or even better pull requests. https://github.com/Na1w/marmel Some demonstration videos of projects this orchestrator completed successfully fully autonomously (althought previous versions). NES and C64 emulation https://www.youtube.com/watch?v=XYjwAFk6wM0 Full "live coding" example of a connect 4 game (also demonstrates the philosophy behind the tool) https://www.youtube.com/watch?v=Mj_CwnI0f98