fix: respect user round count in deep research (#2896)

The STOP_PROMPT did not include the target round count, so the LLM
could decide to stop after 2-3 rounds even when the user requested 8.
Additionally, min_rounds was capped at 3 regardless of max_rounds.

- Add max_rounds to STOP_PROMPT so the LLM knows the target
- Change min_rounds from min(3, max_rounds) to max(2, max_rounds - 2)

Fixes #2863

Co-authored-by: michaelxer <michaelxer@users.noreply.github.com>

src/deep_research.py

@@ -107,7 +107,7 @@ You are deciding whether a research report is comprehensive enough.
 **Current report:**
 {report}
 
-**Rounds completed:** {round_num}
+**Rounds completed:** {round_num} of {max_rounds}
 
 Based on the report so far, do we have enough information to answer the question \
 comprehensively?  Consider:
@@ -115,6 +115,9 @@ comprehensively?  Consider:
 - Are there obvious gaps or unanswered sub-questions?
 - Is the evidence sufficient and from multiple sources?
 
+If rounds completed is well below the target, prefer continuing unless the \
+report is already exhaustive.
+
 Reply with ONLY "YES" or "NO" followed by a brief one-sentence reason.
 Example: "YES — The report covers all major aspects with evidence from multiple sources."
 Example: "NO — We still lack information about the economic impact."
@@ -698,6 +701,7 @@ class DeepResearcher:
             question=question,
             report=report,
             round_num=round_num,
+            max_rounds=self.max_rounds,
         )
 
         try:

src/research_handler.py

@@ -777,7 +777,7 @@ class ResearchHandler:
                 llm_model=llm_model,
                 llm_headers=llm_headers,
                 max_rounds=max_rounds,
-                min_rounds=min(3, max_rounds),
+                min_rounds=max(2, max_rounds - 2),
                 max_time=max_time,
                 max_report_tokens=_max_report_tokens,
                 extraction_timeout=_extraction_timeout,