Optimizing tool-use by guiding without explicit instruction

[ErikBjare]

Optimizing tool-use messages like saves and patches by rewriting inefficient calls to more efficient versions, guiding the model towards correct tool-use without explicit instruction/self-correction.

Since a while back we emit a warning if the agent made an inefficient patch, the agent will respect the warning, but also annoying acknowledge it. I've thought of adding the reverse for save (including better save previews), but instead I got this idea of quietly replacing inefficient save/patches with efficient versions, subtly guiding the LLM to the better way.

Saving:

• Inefficient saves could be replaced by patch

Patches:

• Verbose patches (excessive context) could be reduced
• Inefficient patches could be replaced by save

I'm not sure if there are other cases where we could/should do the same.

We could also use such a rewriting strategy to replace invalid save/patch formats with valid ones: #153