Hey Prime. I work on a similar, in-between, tradcoding and vibecoding approach to AI coding robertpiosik/CodeWebChat. Cool to see you're researching non-agentic tooling. I'd love to share everything I learned building CWC but let's focus on the most important thing now, that is picking context files.

What can't be done by an agent and what have an immense influence on the way you prompt the model and its accuracy is including the right modules and examples from the start. I call it model guidance through context ownership.

What you did with inline completions could be useful but is almost never a desirable way to prompt the model as it can't modify anything outside the function body, e.g. can't add properties to a class or a constant in a separate file.