LLM model selection per role

[baaelke]

Summary

Currently all agent roles (planner, coder, reviewer, summarizer) share a single model config agent.code.model. This makes it impossible to use e.g. a cheap model for reviewing and a powerful one for coding.

Proposed config

[agent.planner]
model = "gpt-5.4"
model_temp = 1.0

[agent.coder]
model = "gpt-5.4"
model_temp = 1.0

[agent.reviewer]
model = "gpt-5-mini"
model_temp = 0.3

[agent.summarizer]
model = "gpt-5-nano"
model_temp = 0.0

Tasks

• Add specific role configs to config.py and AgentConfig
• Update Query.__init__ to accept a role parameter and read the matching config
• Assign correct role to all Query() calls in minimal_agent.py
• Extend /model command in main.py to support per-role switching (/model planner gpt-5-nano) (see feat: add /model command #43 for /model command)
• Update default config.toml
• Maybe keep agent.code.model as fallback

Affected files

config.py, config.toml, treesearch/llm/query.py, treesearch/minimal_agent.py, main.py

[eisenbahnhero]

That doesn't sound like a bad idea to me. Do you already have any initial findings or ideas about how much this affects the cost of a run or the results?