I built a coding harness on top of the new Opencode v2, it started with the the idea of using the new jev model for routing easy tasks to cheaper models such as deepseek and hard tasks to models such as Opus 5.5. In total in separated it into 4 rungs - easy, medium, hard and frontier.
Two more layers compound the savings: the Sol-Pi methodology based on the recently published paper which uses action fusion and observation packing and cuts roken usage within every tier. Doctrine injection loads your engineering standards into every session by default, so the agent writes code the way your team would.
There is also a real control plane - escalation, permission and verification hooks so the agent surfaces uncertainty instead of being confidently wrong in silence.
Run it using your own key from Openrouter or Vercel AI gateway key, we need this since one key is able to call the jev decision model as well as the llms for coding.
One more thing I added was the ability to connect your Meta Muse to use this for coding tasks which has worked really well so far you can try it out by copying the promp for agent on the website.
Two ways you can use it today is the Mac app and the cli.
Honest limitations are documented, happy to take any questions.
Repo: [https://github.com/grist-ai/grist](https://github.com/grist-ai/grist)
npm CLI: npm install -g grist-ai
Comments URL: [https://news.ycombinator.com/item?id=49928275](https://news.ycombinator.com/item?id=49928275)
Points: 1