[FEATURE] promptSuggestionEnabled defaults to true and inherits the session's model — should default to false, or use a fixed low-cost model like Haiku (same pattern already used for session-title generation)

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

The inline prompt suggestion feature (grayed-out suggestion text in the input box, controlled by promptSuggestionEnabled in settings.json) is enabled by default. Based on network traffic inspection, each suggestion call inherits the model of the current session rather than routing through a fixed low-cost model.

Proposed Solution

1. Default promptSuggestionEnabled to false, making it an explicit opt-in. Users who want the convenience can turn it on knowingly; users who don't need it stop paying for it silently.
2. If it stays on by default, decouple it from the session's model and route it through a fixed low-cost model (e.g. Haiku), matching the existing approach used for session-title generation.

Either change removes the current situation where a cosmetic input-box feature can silently consume premium-model quota without the user ever choosing that trade-off.

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

Other

Use Case Example

_No response_

Additional Context

Steps to reproduce / how I verified this

1. Start a Claude Code session using a high-cost model (e.g. Fable 5).
2. Begin typing in the input box and pause partway through a word/phrase so the inline suggestion triggers.
3. Capture the network request for that suggestion call.
4. Observe that the request targets the same model as the active session, not a fixed cheaper model.

As a control, I also captured the network request for automatic session-title generation, which is consistently routed to Haiku regardless of the session's active model. This confirms the pattern of isolating a lightweight, non-core feature from the main model choice already exists in the codebase — it just isn't applied to prompt suggestions.

Why this matters

• Prompt suggestion is a convenience feature, not core functionality. Many users don't need it and would be fine without it entirely.
• Because it inherits the session model, the cost is unpredictable from the user's side: the same feature is nearly free in one session and noticeably costly in another, purely based on which model happens to be active — a distinction most users won't think to check before typing.
• The input-side cost is also opaque. Even early in a session with minimal context, users have no visibility into how much context is sent with each suggestion call. Combined with an inherited premium-rate model, this compounds an already-unclear cost.
• This is inconsistent with how Anthropic already handles a comparable lightweight feature (session-title generation locked to Haiku), suggesting the fix is a matter of applying an existing pattern rather than building something new.