Force clarifier agent to trigger searches before asking questions

[cdgamarose-nv]

If the clarifier LLM doesn't know about something from its internal knowledge (training data), it immediately asks questions. Since we have a clarifier agent with a tool-calling loop, we need to trigger searches to gather more context before asking questions.

Example:

Expected

This should have triggered a web search to find out context and then ask more questions if needed. This should not be overly prompt engineered for one model but instead should work with most frontier and open models.

[lakshaymor4]

Hi, @cdgamarose-nv I'd like to work on this. My plan is to add web search tool access to the clarifier agent and update its prompt to search before asking questions. Does this approach sound right?

[torkian]

Took a crack at this in #245. Approach summary:

• Structural fix at the LangGraph level (not prompt-only, per the issue's "should not be overly prompt engineered" guidance): a new force_search node nudges the clarifier to call its bound search tools before falling through to ask the user.
• Single-use guard (force_search_used flag) so a model that ignores the nudge falls through to ask_for_clarification instead of looping.
• Ephemeral guidance injection (same pattern as JSON_REMINDER_AFTER_TOOLS) so internal scaffolding can't leak into user-facing fallback text.
• Supporting prompt update so frontier models do the right thing voluntarily and skip the nudge entirely.

Verified end-to-end against nvidia/nemotron-3-nano-30b-a3b (the repo's default clarifier model) with a real Tavily web search:

• develop: 0 tool calls, user prompted immediately (reproduces the bug)
• this branch: 1 tool call, user never prompted

148 clarifier tests pass (10 new). Happy to address review feedback.