Separate explorer agent

[aa956]

Just an idea, read the reddit post today:

https://www.reddit.com/r/LocalLLaMA/comments/1ud1lro/why_is_no_one_talking_about_microsofts_open/

And checked the HF model repository for Qwen3-4B-Instruct based repository-exploration subagent model:

https://huggingface.co/microsoft/FastContext-1.0-4B-SFT

And I remember that as soon as project grows over a few source files a lot of architect context may go into initial exploration.

Could be improved with good documentation for the agent (e.g. AGENTS.md/README.md describing project files, conventions, code location) but also probably could be solved by something like the separate agent that is not a coder but explorer?

Especially if late-cli is started to check new project/repository for the first time.

So instead of architect trying this itself:

find /project -type f -name "*.md" -o -name "*.py" -o -name "*.yaml" -o -name "*.yml" -o -name "*.json" -o -name "*.toml" | head -80
ls -la /project
#read_file /project/AGENTS.md
#read_file /project/README.md
# Good that model checks first 30, we've 8Gb .venv with pytorch + cuda
find /project -type f -name "*.py" | head -30 
find /project -type f -name "*.py" -not -path "*/.venv/*" | head -30
#read_file /project/src/main.py
#read_file /project/src/module/module.py
#...

It can ask subagent to do it?

Maybe only after e.g. /README.md and /AGENTS.md did not give clear enough project overview (that can be verified by architect to not be obsolete with a single ls call).

[mlhher]

From reading the MS repo it seems it is basically just using Lates architecture but instead of for writing code it does it for searching the code.

I think with #83 this might be a good next step; I was generally thinking about subagents to add so this might be the best next one that provides immediate return.
When #83 is merged I will look into this.

Thank you for the suggestion!