Mode-aware reasoning effort / model switching (plan mode vs execution)

Problem

Plan mode benefits from high reasoning effort (deep thinking, architecture, trade-offs), while execution mode benefits from faster, cheaper inference. Today there's no way to switch automatically — MAX_THINKING_TOKENS and model selection in settings.json are global and require a session restart.

Users can manually flip profiles (e.g. via a keybinding that swaps Opus+high-thinking ↔ Sonnet+medium), but this relies on remembering to toggle on every mode change.

Requested

One of:

1. Mode-aware settings — allow settings.json to specify different model / MAX_THINKING_TOKENS per permission mode:

``json
{
"modes": {
"plan": { "model": "opus", "maxThinkingTokens": 32000 },
"acceptEdits": { "model": "sonnet", "maxThinkingTokens": 4000 }
}
}
``

1. Mode exposed to hooks — surface the current permission mode in UserPromptSubmit / add a ModeChange hook event, plus a supported way for a hook to adjust model / thinking budget for the next turn.

Either unlocks the pattern "think hard while planning, move fast while executing" without manual toggling.

Why it matters

• Aligns cost/latency with the actual cognitive demand of each phase
• Removes a class of "oops, forgot to switch back to Sonnet" waste
• Plays well with the existing plan-mode UX — the signal is already there, it's just not actionable