When I noticed the apparent return to our previous state, I pointed this out, and Gpt 5 quickly escaped the loop with a new code sample that fixed two different problems across a two-day span of discussions. It did this with only small code samples spread across the chat history. Very impressive.

I'm developing the feature live on twitch.tv/CodeRushed at the top of the hour -- working on UI today.

Today was the first day I used the feature in a non-test environment -- because I needed it -- on real code, and the results were surprisingly impressive. The code changes were correct and folded in right where I needed them. The work is still at an early stage, but the possibilities are exciting.