Rework --llm-config CLI arg

[li-boxuan]

What is the problem that this fixes or functionality that this introduces? Does it fix any open issues?

After #2756, the semantics of --llm-config CLI arg becomes ambiguous: what exactly does it mean to provide a LLM config group via CLI, especially when agent delegation is involved?

---

Give a summary of what the PR does, explaining any non-trivial design decisions

This PR makes it clear that --llm-config allows one to override default LLM config ([llm] section in config.toml) with a given config group. For example, say you have

[llm.eval_gpt4_1106_preview_llm]
model = "gpt-4-1106-preview"
api_key = "XXX"
temperature = 0.0

[llm.eval_some_openai_compatible_model_llm]
model = "openai/MODEL_NAME"
base_url = "https://OPENAI_COMPATIBLE_URL/v1"
api_key = "XXX"
temperature = 0.0

Then you can use --llm-config eval_gpt4_1106_preview_llm or --llm-config eval_some_openai_compatible_model_llm to control the default LLM config used in the entire lifecycle of main.py.

What if I also have the following in my config.yaml?

[agent.BrowsingAgent]
llm_config="gpt-4o"

Then BrowsingAgent would use llm.gpt-4o while other agents would use --llm-config.

[li-boxuan]

Testing in progress...

Done. Working as expected.

[xingyaoww]

Thanks! LGTM!

[xingyaoww]

Does it mean that, if I still specify the model config for a particular agent (e.g., browsing), while also override default config via --llm-config - It will use the overrided default config (--llm-config) for other agents, and will use the agent-specific config in config.toml for browsing agent?

[li-boxuan]

Yep