Proposal: follow-up question suggestion chips under assistant responses

[mvanhorn]

Perplexity, NotebookLM, and ChatGPT all render 3 clickable follow-up question chips under each response. AnythingLLM, Khoj, and Open-Notebook don't. SurfSense's system_prompt.py already acknowledges follow-ups as a supported message type (line 168) but doesn't surface suggestions.

I'd add a suggest_followups tool that runs as the last step of each response and emits [string, string, string]. Frontend renders them as <Button variant="secondary" size="sm"> chips below the existing AssistantActionBar. Click → submits as a new user message via useThreadRuntime().append(). Feature-flagged via ENABLE_FOLLOWUP_SUGGESTIONS env (default on).

Scope: ~140 LOC total. Small, additive, no existing behavior change when flag is off.

Two quick questions:

1. Tool vs. appending to main response: I'd use a separate tool to keep the main response format clean — agree, or do you prefer structured suggestions within the response envelope?
2. Character budget: I'm targeting ≤12 words per chip. Acceptable or want stricter/looser?

I've shipped 3 PRs on SurfSense (#1215, #1220, #1233). Targeting dev, full unit test on the tool output schema, screenshot + GIF in the PR body.