Bring back auto-scaling of effort level / model

Problem

Fixed effort level and fixed model means every turn is treated the same:

• Simple turns (push a branch, file an issue, acknowledge a command) run at whatever high setting is configured — wasteful tokens, slow response.
• Complex turns (architecture decisions, multi-agent orchestration, deep design debates) can't scale above the configured level either.

A multi-turn session like orchestrating a feature-workflow PR mixes both kinds constantly. Right-sizing per turn currently requires manual /effort, /model, or /fast toggles on every transition, which no one actually does — so the default dominates and is wrong for most turns.

What changed

Earlier versions had auto-scaling of effort/model based on task shape. Current behavior is strictly fixed by the configured effortLevel and model settings until the user manually overrides.

Request

Restore auto-determination of thinking effort and model selection per turn based on task complexity. The CLI should pay the cost when the task warrants it and save it when the task doesn't.

Current workarounds (all manual, all forgotten during real work)

• /effort per turn
• /model per turn
• /fast per session

Impact

Unusable productivity trade: either slow on trivia and adequate on hard questions (at xhigh), or fast on trivia and inadequate on hard questions (at medium). No setting serves a mixed-complexity session well.