Open
Description
Devon burns excessive amounts of tokens reading a non abbreviated version of the tree. So there either needs to be a light LLM client side which abbreviates it or a always-active request to the user to perform the action, allowing the user to submit their own abbreviated version. With these rate limits, allowing Devon to run known tasks that can eat up a ton of usage is illogical. I see the pair programmers of the future using lite llms on their local machines to reduce the token exposure going up. For the specific task of abbreviating a file tree, you might not even need a full LLM. A simpler natural language processing (NLP) model or even a well-crafted rule-based system could potentially handle this task effectively.
Metadata
Metadata
Assignees
Labels
No labels