Auto mode: classifier that picks model (Haiku/Sonnet/Opus) and thinking effort per task

Problem
There are now 3 model tiers × 5 thinking-effort levels = 15 combinations. I don't have good intuition for when Sonnet-medium beats Opus-low, or when Haiku is enough. In practice I either over-spend (Opus for everything) or under-spend (Sonnet for something that needed Opus and then re-do it). The mental tax of picking the right config every turn is real.

Proposal
An "auto" mode that classifies the incoming task (complexity, breadth, novelty, safety stakes) and picks the model + thinking level for that turn. Show the chosen config in the UI so I can learn from it and override when wrong. Ideally the classifier is itself a cheap Haiku call so it doesn't add meaningful latency/cost.

Why it matters
Right now the "which model" decision is a friction tax on every non-trivial turn. Auto mode would remove that tax and also give new users a sensible default instead of making them read model comparison docs.