I assume we have a very different workflow.
I'm using VSC for most edits, tab-completion is done via Copilot, I don't use it that much though, as I find the prediction to be subpar or too wordy in case of commenting. I use Aider for rubber-ducking and implementing small to mid-scope changes. Normally, I add the required files, change to architect or ask mode (depends on the problem I want to solve), explain what my problem is and how I want it to be solved. If the Aider answer satisfies me, I change to coding mode and allow the changes.
No magic, I have no idea how a single prompt can generate $4. I wouldn't be surprised if I'm only scratching on the surface with my approach though, maybe there is a better but more costly strategy yielding better results which I just didn't realize yet.
Probably related to Sonnet 3.7’s rampant ADHD and less the CLI tool itself (and maybe a bit of LLMs-suck-at-Swift?)
With deepseek: ~nothing.
My only problem was deepseek occasionally not answering at all, but generally it was fast (non thinking that was).
You your /tokens to see how many tokens it has in its context for the next request. You manage it by dropping files and clearing the context.