Suggest the next prompt after the agent finishes #21429

Problem

When an agent finishes a turn, the next message is often predictable: approve the plan it proposed, answer the question it asked, or ask it to verify, commit, or open a PR. Several agent-first tools now suggest that next prompt. In T3 Code you type it every time.

Proposal

An opt-in setting, Settings > General > Next prompt suggestions, off by default.

• After a turn you watched finishes, the empty composer shows the most likely next message as its placeholder. Tab fills it in, so you can edit it or send it. Typing anything else ignores it.
• The suggestion comes from a text generation model, using the agent's latest reply and your last few messages, written in your voice. By default it uses the existing text generation model. An optional Prompt suggestion model lets you pick a fast, cheap model just for this, the same way review guides can override the writer model.
• Cost and noise are bounded. It only runs for replies that finish while the thread is open; opening or switching to old threads never runs a model. Rolled-back turns are excluded. Failures fall back to the normal placeholder.

Shape

• One guarded RPC, orchestration.suggestNextPrompt { threadId, afterMessageId }, with a thin PromptSuggestionService behind it.
• One new operation on ProviderTextGeneration; runner-based providers get it from fromRunner.
• Two server settings. No new persisted state, events or projections.

There's a working implementation, with focused tests, in #21428. I'm happy to reshape or narrow it based on feedback, for example dropping the dedicated model setting, or adding the mobile composer.

Questions for maintainers:

1. Is this direction welcome as an opt-in setting?
2. Is placeholder text plus Tab the right interaction, or would you prefer a chip above the composer?