Support better understanding of subagents/teams

[jmaver]

Not quite sure how to suggest a solution for this, but here is the situation:

claude code config:

**    "env": {
        "ANTHROPIC_BASE_URL": "http://192.168.200.83:8666",
        "ANTHROPIC_AUTH_TOKEN": "XXX”,
        "ANTHROPIC_MODEL": "Qwen3.5-35B-A3B-4bit",
        "ANTHROPIC_DEFAULT_HAIKU_MODEL": "Qwen3.5-4B-MLX-4bit",
        "ANTHROPIC_DEFAULT_SONNET_MODEL": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit",
        "ANTHROPIC_DEFAULT_OPUS_MODEL": "Qwen3.5-35B-A3B-4bit",        
        "API_TIMEOUT_MS": "3000000",
        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
        "ENABLE_LSP_TOOL": "1",
        "CLAUDE_CODE_ATTRIBUTION_HEADER":"0"
    },

run claude code
claude --settings settings-omlx.json —worktree

claude starts with default model (35b-a3b)

tell claude to spin up a team of haiku agents and implement something as a test
create a team of 2 haiku agents to implement a simple calculator command line app. one in python and one in iavascript. The calculator only needs to support addition. Each version should be in its own subdirectory and have 5 unit tests and simple documentation

claude runs and spins up subagents, but

1. I don’t know what models are being used in the ui. I can only flip back and forth between them in the status page dropdown and see that the 4b and 35b both had info for this session, but I don’t know what is running. the models page only ever showed the a3b one.
2. I don’t know if omlx loaded both sets of models at once (35B-a3b and 4b) or swapped them back and forth

Anyway, a weird usecase, but subagents and teams will be more common over time and it would be nice if there were a way to see what was running and know more about how they are loaded/unloaded.

Regardless, the experiment worked and omlx made that happen - lmstudio would have choked on this. :)

[jundot]

Being able to see which model is active is technically doable, but i'm not sure if building a full GUI for it in the admin panel is the right call resource-wise. It would look nice and be somewhat convenient, but for immediate needs, checking the logs should tell you what's going on.

I just realized the logs don't actually show model names right now either. I'll fix that. And i'll think a bit more about the best way to surface which models are currently using resources.