Improve thinking prompt system with better constraints and history handling

KiroProxy/kiro_proxy/core/thinking.py

@@ -318,28 +318,59 @@ def format_thinking_block(thinking_content: str) -> str:
     return f"<thinking>\n{thinking_content}\n</thinking>"
 
 
-def build_thinking_prompt(user_content: str, *, budget_tokens: Optional[int]) -> str:
-    """Build a separate prompt using Tree of Thoughts approach.
+def build_thinking_prompt(
+    user_content: str, 
+    *, 
+    budget_tokens: Optional[int],
+    history: list = None,
+    has_tool_results: bool = False
+) -> str:
+    """Build a thinking prompt for internal reasoning phase.
     
-    Use multiple expert perspectives to analyze the problem deeply.
+    Key: This is planning-only, no actual tool execution.
+    The model will receive full conversation history via the API.
     """
-    if user_content is None:
-        user_content = ""
-
     budget_str = ""
     if budget_tokens:
-        budget_str = f" Budget: {budget_tokens} tokens."
+        budget_str = f" (Budget: {budget_tokens} tokens)"
+    
+    # 根据是否有工具结果调整提示
+    if has_tool_results:
+        return (
+            f"[INTERNAL REASONING PHASE]{budget_str}\n\n"
+            "You have received tool execution results. Based on the COMPLETE conversation history above, "
+            "think through what these results mean and plan your response.\n\n"
+            "CRITICAL CONSTRAINTS:\n"
+            "- DO NOT call any tools in this thinking phase - analysis only\n"
+            "- DO NOT output JSON tool calls or function_call structures\n"
+            "- Output ONLY plain text reasoning\n"
+            "- IMPORTANT: Think in the SAME LANGUAGE as the user's request\n\n"
+            "Think about:\n"
+            "1. What was the user's original request (from conversation history)?\n"
+            "2. What do the tool results tell us?\n"
+            "3. Is the task complete, or are more steps needed?\n"
+            "4. How should I synthesize this into a helpful response?\n\n"
+            f"Current input: {user_content}"
+        )
     
     return (
-        f"Think deeply and comprehensively about this problem.{budget_str}\n\n"
-        "Use the following approach:\n"
+        f"[INTERNAL REASONING PHASE]{budget_str}\n\n"
+        "Based on the conversation history above (if any), analyze the current request thoroughly.\n\n"
+        "CRITICAL CONSTRAINTS:\n"
+        "- DO NOT call any tools - planning and analysis only\n"
+        "- DO NOT output JSON tool calls, function_call, or tool_use structures\n"
+        "- Output ONLY plain text reasoning\n"
+        "- IMPORTANT: Think in the SAME LANGUAGE as the user's request\n\n"
+        "Think about:\n"
         "1. Break down the problem into components\n"
-        "2. Consider multiple perspectives and solutions\n"
+        "2. Consider multiple approaches and perspectives\n"
         "3. Evaluate trade-offs and edge cases\n"
-        "4. Synthesize your analysis into a coherent response\n\n"
-        f"{user_content}"
+        "4. Plan what actions/tools would be needed (without executing)\n"
+        "5. Synthesize into a coherent strategy\n\n"
+        f"Current request: {user_content}"
     )
 
+
 def build_user_prompt_with_thinking(user_content: str, thinking_content: str) -> str:
     """Inject thinking into the main prompt.
     
@@ -402,8 +433,14 @@ async def fetch_thinking_text(
     timeout_s: float = 600.0,
 ) -> str:
     """Non-streaming helper to get thinking content (best-effort)."""
-    thinking_prompt = build_thinking_prompt(user_content, budget_tokens=budget_tokens)
     clean_history = strip_thinking_from_history(history)
+    has_tool_results = bool(tool_results)
+    thinking_prompt = build_thinking_prompt(
+        user_content, 
+        budget_tokens=budget_tokens, 
+        history=clean_history,
+        has_tool_results=has_tool_results
+    )
     thinking_request = build_kiro_request(
         thinking_prompt,
         model,
@@ -435,8 +472,14 @@ async def stream_thinking_text(
     timeout_s: float = 600.0,
 ) -> AsyncIterator[str]:
     """Streaming helper to yield thinking content incrementally (best-effort)."""
-    thinking_prompt = build_thinking_prompt(user_content, budget_tokens=budget_tokens)
     clean_history = strip_thinking_from_history(history)
+    has_tool_results = bool(tool_results)
+    thinking_prompt = build_thinking_prompt(
+        user_content,
+        budget_tokens=budget_tokens,
+        history=clean_history,
+        has_tool_results=has_tool_results
+    )
     thinking_request = build_kiro_request(
         thinking_prompt,
         model,