unphased
commented
5 days ago I agree totally with this concept. It seems to already exist as we already get prompted by aider to do things like add files that the LLM suggests it wants access to, and for the LLM to add new files and such things. I have more applications of this concept to add to the discussion as well.
One very big bottleneck that high level LLM tools like aider address very well is the issue of repetitively copy-pasting code. Now we do have to repeat our code in the prompts a significant amount, because that's what's required to update the LLM's understanding of changes to the code which we as the primary programmer are making. Sometimes these updates may be compressed with descriptions or diffs, but just like with math, an LLM is poor at tracking mutated state.
As we blow past this bottleneck, then, the tradeoff is API token consumption which costs directly scale with. When we do the caveman copy paste workflow, we manually manipulate text and get a deeply intuitive sense of token consumption. Granted, with chatbot usage the pricing isn't per-token, but we need to manage token consumption for the purpose of keeping important information in-context anyhow, as results drop off a cliff when we run afoul of that.
What I'm trying to get at is with an open-source tool like aider we have a unique opportunity to give excellent UX that no AI vendor will ever offer, that is to say something designed to help the user control resource consumption.
So I would suggest for Review Mode to expand its scope beyond just being an approval process for changes: I want to apply this approval process also to prompts whose token consumption exceed a certain threshold (heuristically or otherwise) as one method to help us keep token related costs in check.
A lite version of this could be some sort of "more raw" debug log to gain access to where we can view all of the raw tokens being sent in and received from the LLM. Even if no Review Mode approval interaction exists we could use a raw communication log to at least see how the tool's automation works so we can learn how to better control it.
One thing i'm not clear on is whether aider could automatically endeavor to autorespond with increased contextual detail if it becomes obvious somehow that not enough context was provided initially. If this were to happen, it may lead to greatly increased or even runaway token consumption.
Edit: OK I just came across https://github.com/paul-gauthier/aider/issues/127#issuecomment-1641587378 and it led me to discover the verbose flag, which is very useful in this regard as it appears to be doing exactly what my "lite version" concept above is, letting me see exactly what's being sent and received so I can gauge if I'm being efficient in token usage. the /tokens command is also a super useful thing to help us predict the cost of a next instruction we may send.
I'm REALLY impressed with this tool right now. I might suggest two minor quality of life things related to verbose:
- A way to get that verbose content but shoved in a file not unlike
.aider.chat.history.mdso i can ignore it most of the time but review or tail it as desired, without it cluttering the main terminal output - A mode that spits out a summary of the real incurred costs after a command completes executing so I don't have to be quite as glued to the billing console of the AI vendor
paul-gauthier
Context: When I ask Aider to review my code, it often identifies issues and suggests corrections. While some of these corrections are accurate, others are based on design choices that do not need to be changed. Currently, Aider goes ahead and implements all suggested changes and commits them, which sometimes results in unnecessary modifications that I need to undo, leading to diminishing returns.
Request: I propose adding a "Review Mode" switch in the Aider browser interface, with an accompanying checklist feature. This mode would allow users to control how Aider behaves when reviewing code. In Review Mode, Aider would:
Workflow:
Benefits:
Implementing this feature would streamline the code review process, ensuring that changes made by Aider align more closely with the user's intentions and providing a more interactive and user-friendly experience.
Aider v0.40.0 Models: claude-3-5-sonnet-20240620 with diff edit format, weak model claude-3-haiku-20240307 Git repo: \192.168.0.35\YonaVmDataShare\Projects\posexport.git with 28 files Repo-map: using 1024 tokens Restored previous conversation history.