[FEATURE] Selectable ghost-text prompt suggestions — cycle through N, not just one

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

Prompt suggestions offer a single greyed-out guess at my next message, accepted with Tab / Right-arrow. When the guess is right it saves real typing. When it's wrong — which is often — the feature contributes nothing and I type my reply out by hand.

The issue is that my next message is usually one of a small predictable set, not one specific string.

A concrete case from a long debugging session. Claude finished a diagnosis and ended on a recommendation. My realistic next messages were:

• fix it
• explain more first
• skip it, move on

Claude can predict that set at least as well as it predicts any single member of it. Today I get one guess out of three, so most of the time the suggestion is simply the wrong one and I get no benefit.

The value scales with how long Claude's message was. After a wall of analysis, my reply is often a two-word directive — which is exactly where typing feels most disproportionate to the size of the decision I'm actually making.

Proposed Solution

Keep the current interaction exactly as-is and add cycling:

> fix it                              ← ghost text, suggestion 1/3
  [Tab] next  ·  [→] accept  ·  type to dismiss

> explain more first                  ← Tab once, suggestion 2/3
> skip it, move on                    ← Tab twice, suggestion 3/3

• Right-arrow still accepts whatever is currently displayed — existing muscle memory untouched
• Tab (or Down/Up) cycles between suggestions
• Typing still dismisses
• A subtle 1/3 indicator so it's discoverable that more exist

Nothing changes for anyone who ignores it, and if only one suggestion is generated the behavior is identical to today. This is additive.

Alternative Solutions

A popup list below the prompt (shell-completion style, arrows + Enter). More immediately discoverable, but more visually invasive and it competes with existing terminal completion UI. I'd prefer the cycling model above, which costs zero screen space.

Current workaround: none. If the single suggestion isn't the one I want, I type the message manually. There's no way to ask for another guess or see alternatives.

Priority

Medium - Would be very helpful

Feature Category

Interactive mode (TUI)

Use Case Example

1. I ask Claude to diagnose a failing integration in a large codebase.
2. Claude investigates across several files and returns a long analysis ending with a recommended fix.
3. I read it and agree. My reply is literally fix it.
4. Today: the ghost text guesses something else, so I type fix it by hand. The suggestion cost a background request and delivered nothing.
5. With this feature: fix it / explain more first / skip it, move on are all one Tab-cycle away. One keystroke instead of a sentence.

This repeats many times per session. Long analysis → short directive is the dominant rhythm of an agentic coding session, and it's precisely the case where a single guess has the worst hit rate: the model knows the shape of my likely response but has to gamble on which one.

Additional Context

This capability already exists in the product. Claude Code has AskUserQuestion, which presents 2–4 selectable options — and it's excellent. I hit it several times in the session that prompted this request, and it was consistently the smoothest interaction of the day.

But it's model-invoked: it only appears when Claude decides to ask a question. I can't request it, can't express a preference for it, and it's entirely disconnected from the passive ghost-text suggestion path.

So both halves already ship — the model can generate good discrete options, and there's UI for selecting between them. This request is largely about wiring an existing capability into the passive suggestion path, rather than building something new.

Related but distinct issues (deliberately not folded in here, to keep this a single request):

• #74826 — frequency / always-on mode for prompt suggestions. Orthogonal: that's about how often suggestions appear, this is about how many are offered when they do.
• #76591 — inline Tab hint for accepting ghost text. Complementary: the 1/3 indicator proposed above would want to solve discoverability the same way, and the two should probably share a design.
• #72495 — suggestions suppressed near rate limits. Unrelated cause, but worth noting these three plus this request all circle the same feature getting under-used relative to how good it is when it lands.