:CodeCompanion output is inlcuding the backtick markdown code fences

[mrjones2014]

Is this something wrong with my config, or do we need to update the system prompt to tell the model not to include the triple-backticks?

[olimorris]

I think the backticks may need to be explicitly called out. The prompt does ask for no codeblocks but some models have been super resistant to that. GPT-4 went through a phase of completely ignoring it before reverting back to normal.

The init.lua system prompt, line 27, is the one to adjust. Can you see what works for you and then we can include it.

I really don't want to get to having different prompts for different models.

[mrjones2014]

Gotcha, it seems like llama2 is just very resistant to it. I would probably recommend dolphin-mixtral as the default model for code, it's much better than llama2. Part of its system prompt is to follow instructions absolutely no matter what, its #1 priority is following the users instructions, which helps with this kind of thing.

[olimorris]

Good shout. I've now made that the default.