Thinking UX does not preserve full underlying model capabilities #281
Description
A key design principle of cmux is we give users the full power of the underlying model. Our Thinking slider violates that.
Currently, we have a slider with four levels: off, low, medium, high
This maps to API as:
| cmux | OpenAI | Anthropic |
|---|---|---|
| off | off | 0 tokens |
| N/A | minimal | N/A |
| low | low | 4000 tokens |
| medium | medium | 10000 tokens |
| high | high | 20000 tokens |
We miss the "minimal" level by OpenAI, and the Anthropic token budgets are set arbitrarily. Instead, the Thinking toggle should expose the full set of options for OpenAI and be a token input box for Anthropic. We should establish this pattern of preserving full fidelity on reasoning effort as models advance.