Give more control over "Codebase indexing" process (run indexing when idle)

[kajiev]

This one is important for all the localhost_ers_ IMHO.
I believe, the typical self-hosted solution is:

• ollama
• VRAM utilization 99.9%

Given this setup, when "codebase indexing" wakes up during whatever agent task , this forces Ollama to unload "main" model because it decides to free up some VRAM for embedding model. This results in a kind of race condition between "main" and embedding models. Usually, it ends in trouble.

Suggestion is:

• either perform codebase indexing only when agent is idle;
• and/or add some code_index tool so that the agent could run it explicitly during task execution (to keep index updated).

Anyway, IMHO codebase indexing during uncompleted task often is a waste of tokens, so it is worth posponing it.