- implement some core tools like, read grep
- get a llm provider working - i used openai (5.3-codex)
- implement the core loop
- start with a simple append only outputs in the ui, not a fully functioning TUI

This is enough to bootstrap kon using kon

I think when i reached this stage i started using kon to build kon. Now i only use kon so its been adding/modifying itself for a while now.

I've implemented coding agents in the past but only the core loop and tools, not all the way upto ui, so insights dawn on you that typically won't if you go all the way, like:
- why storing the conversation in json format is easy but eventually a bad idea
- storing ui events along with tool calls, thinking and text blocks to be able to reload them when loading an older conversation (now its obvious to me, but when i was implementing it this bit was a surprise to me)
- the fact that all the openai api compatible implementations could have little quirks that need to be addressed before you truly get them working (took some time to figure out how to get thinking working in gemma 4)

gpt-5.3-codex and gpt-5.4 with the plus plan have been my workhorse. i have a glm lite coding plan as well but they are very unreliable these days.

i'm going to implement subagents next (don't need them for paid models, in fact i hate the overuse of subagents by opus), mainly to use local llms for all the grepping/reading to locate code, aka the "codebase-search" subagent to help save tokens burnt by the larger model and help with weekly limits.