Enhance system prompt flexibility options

[kvokka]

The idea is RooCodeInc/Roo-Code#3773 , kilo codebase facing the same problem (since it's a fork), and I'll be happy to switch from Roo to Kilo and implement it in this repo, coz i maintain my own fork is an overkill.

System prompt really needs some love and the ability to evolve with the help of the community, and this should help with that

ready to pick up the pr (here or in Roo), depending on where I get the idea approval first

[kvokka]

Addition to #396

There was a great idea with 3 (or even 4) steps Orchestrator (rather than 2 steps, which is been used on Boomerang mode and friends), and it might be flying perfectly with different model types (at least to drop something to Gemini flash or even Gemma), but the problem is, that to make it pretty i wanna to deal with this one first.

For my case, the new chat is starting with 22k tokens (at the beginning of the context, which is the most valuable spot), and that is the priority

[HadesArchitect]

At the moment details of the implementation aren't clear from the description (here or in RooCodeInc/Roo-Code#3773)

Could you please clarify how exactly you see it done?

[kvokka]

After a while, I understand that adding support in the templates files will be even better, keeping the existing functionality as it is and just extending the editor.

1. Works only if the user provides the .kilocode/.{slug-name}-system-prompt.mustache file (the templating framework can be discussed, it's just the simplest choice)
2. If such a file is provided, then the default system prompt is empty. All the hassle about system prompt generation is on the user.
3. In the template, we forward the context with all the prompt-related context variables. All the hassle about system prompt generation is on the user
4. User writes his/her system prompt, picking ONLY the things which are related to the mode.

Why:

1. There is plenty of duplication in the current prompt implementation
2. The ability to pick ONLY the desired MCP tools. For example, the user may want to only use the commit tool from the git MCP, but the server will give him the list of 30 tools (and all of that is happening in the most valuable spot - the start of the context)
3. Such thing as a detailed OS version or the current time can be the part of system prompt without MCP or user prompt.
4. Some embedded tools are explained over verbose for my taste, and it was in there for the reason, because it was done a while ago with the thoughts that it might be used by small models too. But the current SOTA models can comprehend shorter versions.

It's an even more advanced technique than provides Roo code. But the implementation should be relatively straightforward