90 minutes of agentic coding chaos

Started simple. “Find Draft on my GitHub and build it.” What followed was 90 minutes of problem after problem.

Missing Metal Toolchain. Corrupt model downloads. Wrong code signing certificate. Permission prompts resetting every single rebuild. Each one got fixed, and a download progress UI got added along the way so the model downloads aren’t just a mystery loading spinner anymore.

The real problem is the live text feature — the thing that’s supposed to show your words on screen while you’re talking. The EOU model downloads fine. It loads fine. It just… doesn’t do anything. Three rounds of debugging. Actor isolation issues. Audio thread problems. The callback never fires. Still broken at the end of the session.

At this point I’m seriously considering scrapping live text entirely. The regular dictation works great — you talk, it transcribes, it pastes. That’s the core feature and it’s solid. The live text is supposed to be a nice-to-have but it’s turning into a time sink that keeps pulling attention away from the stuff that actually matters.

Sometimes the right move is to kill the feature. Red bar.