I decided to try OpenCode after my less than impressive experiments with Goose.

OpenCode worked even worse for me than Goose did. Then I figured out my context windows wasn’t being set, and was probably 2K tokens, which is the Ollama default. 2K tokens take about 15 seconds to burn through when you’re using a coding agent. Maybe less than 15 seconds.

Goose has an option to show me the size of the context window, but I wonder if it was lying to me. I probably don’t care enough to figure this out, maybe later. Probably never.

I changed the default context window to 128K which is the biggest window I can use with the gpt-oss 120B model right now. Other models I can use a way bigger context window. At some point I’m going to configure every model to have the biggest context window that can fit in the memory I have available.

The lesson here is that if you exceed the context window, nothing really works. This isn’t a surprise, I’ve seen horror stories online of people running out of context window and everything falling apart.

So now the code generation seems to be working as one would expect. I can create various simple applications with terrible prompts and I get something that mostly works.