Okay, getting back to Nano Kitty.
Quick catch up if you're just joining the conversation. Three weeks ago, I decided to write a long running agent harness that natively integrates Spec Kitty as a process toolkit. The idea is that it should empower users to run extremely complex long horizon projects with time frames of months, rather than hours.
Spec Kitty handles the planning, while the Nano Kitty harness orchestrates the process. Across time.
If you're happy with Claude or Codex, Opencode or something else, you can use the system as a dedicated sub-agent, where the tasks can potentially outlive any session you're currently working on.
And the idea is that this gives you, the developer, more time to work on the things that are actually fun, while allowing you to concisely delegate your inference power between models and providers which can save you money.
Current state:
It's good, it runs. It meets the spec I crystallized two weeks ago.
But I'm holding off releasing it because the situation has changed.
Things I want to solve before release.
1. Provider is a pain in the butt to set up, so I'm borrowing the system from Openclaw to handle it.
2. DeepSeek Harness dropped. It's such a good paper, and piece of code, that I'm thinking there are pieces of it I want to borrow from the system.
The thinking stack trace for one, is a brilliant idea that feels like it obviously needs to be there. The other is the component system, and reversibility.
3. The MCP system for it is incomplete, and needs an upgrade to handle the new stateless version of the MCP standard that dropped. The new standard also carves out tasks as a core primitive, so I don't need to hand roll a tasks function like I was planning on.
It's going to be another week of dev time using the harness to bootstrap itself.