Coding agent progress:
I switched the runtime layer off llama.cpp and over to Ollama, Ollama Cloud, and OpenAI connectors. That ended up solving a whole pile of problems the actor had been fighting.
To make cloud models work properly, the scheduler had to change. Honestly, those were all good changes. I also listened to a recent interview with the Hermes agent people, which got me rethinking the way I was handling caching in the app. Some of the internal wiring had to be reworked to support all of that. I’m not angry though.
I also went on a side quest and pulled in the document ingestion pipeline I wrote for another project, so it now ships with full Inkbite integration. OCR is optional, which is exactly how I wanted it.
The doctor function had to be rebuilt, but the health checks are working. Unit tests are testing. Some of the Bubble Tea pieces had to be ripped out, rearranged, and rethought.
All told, I managed to add five more days to the project, which means I’m now 21 days behind my original schedule. It’s okay, Samuel will forgive me when he sees how awesome this thing is.
That said, I’m very close to having a fully functional, role-based coding actor with memory, file utilities, and sub-agents that do not give you migraines or need much babysitting.
Meanwhile, I’m also thinking about changing project law inside the app so it’s easier to switch development styles. Right now it’s built around a strict TDD-heavy Robot Kanban with aggressive planning discipline, but there’s no reason it has to stay locked that way.
At the moment, I’m feeling pretty damned satisfied with myself.
I also know that feeling is probably temporary.
Functional evals are next.