Can you share more on how your setup has changed over time for running these? Do you prompt them for code samples like some people do with ChatGPT or did you integrate them into your IDE or some kind of custom harness?
Sure. I usually work with devcontainers from vscode. They provide great integration ootb (port forwarding & stuff) and are ok for containing the agents for most cases. If you want to work on docker projects I also tried vagrant for vm with docker inside, and you instantiate the agent from vagrant.
For local models I used mainly cline and then roo code extensions. Roo was a bit better because it offered more customisation (prompts, tool choice, etc). I found that local models need shorter prompts and less tools to be effective. Unfortunately roo seems to be discontinued, no idea what I'll use after it stops working. Cline works fine for most of the cases ootb, especially if you run inference on a platform that supports good kv caching - I use vLLM.
For subscriptions I use their own harness, as you get the best bang for the buck. For 3rd party subscriptions that don't have their own harness I use opencode (I got a very cheap sub for GLM that I use for exploration and oss projects).
