fix: Improve UX - text comparisons, remove preset bloat, fix tool message errors

- Convert grant comparison from markdown table to cleaner text-based format
- Remove past winners preset questions (not needed)
- Fix tool message error by implementing safe message trimming that preserves tool_call/tool pairs
- Add system prompt instruction to eliminate verbose "I'm going to..." announcements
- Assistant now just calls tools and presents results directly

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

src/analyzer/chat/chat_tools.py

@@ -550,9 +550,14 @@ class ChatTools:
             ),
         ]
 
-        table = ["### Side-by-side", "| Field | A | B |", "|---|---|---|"]
+        # Build text-based comparison instead of table
+        comparison_lines = ["### Grant A: " + getf(A, "title")]
         for name, va, vb in fields:
-            table.append(f"| {name} | {va} | {vb} |")
+            comparison_lines.append(f"**{name}:** {va}")
+
+        comparison_lines.append("\n### Grant B: " + getf(B, "title"))
+        for name, va, vb in fields:
+            comparison_lines.append(f"**{name}:** {vb}")
 
         # ---------------- optional insight ----------------
         context_text = ""
@@ -565,10 +570,10 @@ class ChatTools:
         if self.client and self.client.is_ready() and context_text.strip():
             try:
                 prompt = (
-                    "Given the factual table and context below, write 3-5 bullet points "
+                    "Given the factual comparison and context below, write 3-5 bullet points "
                     "highlighting *meaningful differences* that matter to SMEs (funding size, "
                     "duration, eligibility, etc.). Do not restate identical facts.\n\n"
-                    + "\n".join(table)
+                    + "\n".join(comparison_lines)
                     + "\n\n"
                     + context_text
                 )
@@ -578,8 +583,7 @@ class ChatTools:
 
         md = [
             "### Comparison",
-            f"**{getf(A, 'title')}** _(A)_ vs **{getf(B, 'title')}** _(B)_",
-            "\n".join(table),
+            "\n".join(comparison_lines),
         ]
         if insight:
             md += ["\n### Key differences", insight]

src/analyzer/chat/demo_app.py

@@ -42,8 +42,6 @@ PRESET_QUESTIONS = {
     "Compare two grants": "Compare competition-2313 and competition-2314",
     "Grant details": "Tell me about competition-2317 in detail",
     "SME funding options": "What grants are available for SMEs with funding over £100k?",
-    "Past winners of open grants": "Of the grants that are currently open, are there any past winners listed as reference?",
-    "Past AI winners": "Show me past winners related to AI projects",
 }
 
 
@@ -169,7 +167,9 @@ class GrantAnalystDemo:
                         "- Avoid fluff, filler, or over-formality\n"
                         "- Skip headings, bullet points, or 'as an AI' disclaimers\n"
                         "- Prefer short, active sentences\n"
-                        "- Be bold, human, and efficient\n\n"
+                        "- Be bold, human, and efficient\n"
+                        "- NEVER announce what you're going to do - just do it and present the results\n"
+                        "- NO phrases like 'I'll search for', 'Let me find', 'Calling the tool' - just call it\n\n"
                     ),
                 }
             ]
@@ -394,6 +394,53 @@ class GrantAnalystDemo:
             logging.error(f"Tool {tool_name} failed: {e}", exc_info=True)
             return {"error": str(e)}
 
+    def _trim_messages_safely(self, messages: List, max_recent: int = 10) -> List:
+        """
+        Trim messages to reduce context size while preserving tool_call/tool pairs.
+
+        Args:
+            messages: Full message history
+            max_recent: Max number of recent messages to keep (excluding system)
+
+        Returns:
+            Trimmed messages list with system message + recent messages
+        """
+        if len(messages) <= max_recent + 1:  # +1 for system message
+            return messages
+
+        # Always keep system message (index 0)
+        system_msg = messages[0] if messages and messages[0].get("role") == "system" else None
+
+        # Get recent messages
+        recent = messages[-(max_recent):]
+
+        # Check if the first recent message is a tool response without its tool_call
+        # If so, find the corresponding assistant message with tool_calls and include it
+        if recent and recent[0].get("role") == "tool":
+            # Find the preceding assistant message with tool_calls
+            tool_call_id = recent[0].get("tool_call_id")
+            if tool_call_id:
+                # Search backwards from the cutoff point
+                cutoff_idx = len(messages) - max_recent
+                for i in range(cutoff_idx - 1, 0, -1):
+                    msg = messages[i]
+                    if msg.get("role") == "assistant" and hasattr(msg, "tool_calls"):
+                        # Found the assistant message with tool_calls, include it
+                        recent = [msg] + recent
+                        break
+                    # Check if it's a dict with tool_calls key
+                    elif msg.get("role") == "assistant" and "tool_calls" in str(msg):
+                        recent = [msg] + recent
+                        break
+
+        # Build final message list
+        result = []
+        if system_msg:
+            result.append(system_msg)
+        result.extend(recent)
+
+        return result
+
     def chat(self, user_message: str, history: List) -> Tuple[str, List]:
         """
         Process a chat message.
@@ -422,7 +469,8 @@ class GrantAnalystDemo:
             llm_start = time.time()
 
             # Limit conversation history (keep system + last 10 messages)
-            messages_to_send = self.messages[:1] + self.messages[-10:]
+            # BUT preserve tool_call/tool pairs to avoid OpenAI API errors
+            messages_to_send = self._trim_messages_safely(self.messages, max_recent=10)
             if len(self.messages) > len(messages_to_send):
                 logger.debug(
                     f"Trimmed conversation from {len(self.messages)} to {len(messages_to_send)} messages"
@@ -491,7 +539,8 @@ class GrantAnalystDemo:
                 final_start = time.time()
 
                 # Limit conversation history to prevent bloat (keep system + last 10 messages)
-                messages_to_send = self.messages[:1] + self.messages[-10:]
+                # BUT preserve tool_call/tool pairs to avoid OpenAI API errors
+                messages_to_send = self._trim_messages_safely(self.messages, max_recent=10)
                 if len(self.messages) > len(messages_to_send):
                     logger.debug(
                         f"Trimmed conversation from {len(self.messages)} to {len(messages_to_send)} messages"