Thank you for the feedback!
I'll definitely look at adding both of those!
4 karma · joined September 30, 2013
I'll definitely look at adding both of those!
It started when taking the train to work and regularly hitting internet dead zones, so I initially downloaded Ollama and OpenCode to see if I could keep building while offline.
The first iteration was an installer that worked out what models I could reasonably run locally (as I wasn’t really sure what my machine would be able to do).
After that, I kept running into cases where an agent would say it had completed a task with a detailed description of the changes it had made but the file hadn’t actually changed.
The project then expanded into an Orchestrator inside OpenCode. It decides how complex a task is and delegates parts of it to subagents, potentially using different local models. Depending on the task, agents can then explore, plan or research something before passing to a coding agent.
The coding agent updates then gets passed to a verifier and complex changes can also go through a separate reviewer.
So, for simple task:
Coder -> Verifier
or for something more complex:
Explorer -> Planner -> Coder -> Verifier -> Reviewer
There’s also a Researcher when more knowledge is needed.
It’s still an experiment as I’m not sure yet how far this approach can push smaller local models.