Coding agent progress:
Another 80 bugs resolved.
Some big ones, some related to emergent architecture and workflow that shouldn't have been there.
Slack, Discord and Teams integrations... in testing.
I feel like those are enough to start with.
Samuel asked for a CWD change to make switching mid project a little simpler. That turned into a substantial rewrite of the project system, but it's better now.
It works the way it does in Codex, except that the planner's now keeping an active log of the projects it's touched during the session, which Codex does not do.
There's a piece of skills law that I need to review this afternoon when I get up. But other than that, we're through the main frontier on it.
Recertifying the models again.
GTPs, Kimi K2.6, and Minimax 2.7 fully tested so far.
Gemma4 still having command channel issues, even after I tried to accommodate for the weirdness of tooling with that model.
Still probably not supported.
That said, the system should support any OpenAI compatible model with agent training, over 20b params with a context window of 256k or better.
That's the minimum spec.
Today, we're going into tui revisions, barring anything huge and unexpected.
The process is pretty buttery.
It's handling complex multi-day assignments now.
It can work and plan at the same time.
Memory system got a big refactor, but it works reliably now.
The Robot Kanban methodology has been formalized and documented.
It's kind of like having a team of little developer agents that work and check each other until projects are done.
I think people are going to like it.