I was reading this page in the docs:
Noticed that when the little yellow warning icon appears, that means its too large to attach to context. Which is good, right, you don’t want to be attaching too much junk.
However, sometimes, I reference directories, just so I can reference the directory, explicitly, and the intent is not exactly to attach everything in that directory to context.
First, if I reference say my main project directory (the git repository directory usually), or say the src/ directory. I would say, always, the intent is not to actually attach all that context. It is just to reference that directory, so the agent knows the general file system “context” that it should be operating within. I think this generally works, but, given the documentation above, I am now not certain.
But it brings up a question. There are context attachemnts…a whole source code file, or lines X–Y of a file. Would it make sense, to add support for more limited “references” that can be included in a prompt or context, that explicitly and only REFERENCE the file or path, without bringing everything inside the file or path into the context (at least, not yet…the model could choose to read more later during the chat if necessary, right?)
So maybe something like: @@src/ could be a contextual reference, whereas @src/ would be a full context attachement (at least, if @src/ and its contents were not too large.)