Weird code merging from deepseek

[viraptor]

I've tried using plandex with deepseek. The responses I see in plandex convo seem completely fine and include the whole file being changed.
Unfortunately plandex is spinning for a very long time processing those answers and finally injects a lot of copied code - the new code gets pasted in the middle of the file being edited. (python code)

I'd love for those edits to be handled better.

[viraptor]

Aider seems to ask for a search/replace pair when working with deepseek and for any issues applying it it just complains to the LLM and makes it try again. Seems to work quite well in practice.

[danenania]

Are you using deepseek for all the roles? You can check with plandex models.

There's a syntax check and a verification step as well, but it may be that deepseek is having trouble with Plandex's line-number based editing. There are pros and cons to this approach vs. aider's search/replace pair strategy... the solution will likely be to change strategies depending on which model is being used, which we are working towards. The prerequisite is a good set of evals that lets us compare different prompting/editing strategies against different models, and we are close to having this in place. So stay tuned :)

[viraptor]

Yup, all the roles. It seems like everything works fine, apart from the change merging.

For the editing format, yes, aider uses diff (as in a search/replace pair) instead for deepseek coder. They already did that homework here so it can save you some time 😉 https://aider.chat/docs/leaderboards/ (includes % reliability for the best method)