[FEAT] - Thinking budgets

[krissetto]

What you'd like to see

cagent should support configuring thinking budgets that can be used for thinking models like gemini-2.5-pro, gpt-5, etc.

Why you'd like to see it

It'd help with tailoring the behavior of those models, where sometimes a user might prefer a low amount of thinking is better because of lower latency, and sometimes a user might want the model to think as much as possible to solve a harder problem

Workarounds?

Not really, just picking different models based on their default behavior

Additional context

e.g.

models:
  low-thinker:
    provider: openai
    model: gpt-5
    thinking_budget: "low"/"medium"/"high"
    ...
  ...

# or we could even support a token based budget, depending on model support
models:
  average-thinker:
    provider: google
    model: gemini-2.5-pro
    thinking_budget: 500 # tokens!
    ...
  non-thinker:
    provider: google
    model: gemini-2.5-pro
    thinking_budget: 0 # tokens!
  ...