hdsp-jupyter-extension 2.0.11__py3-none-any.whl → 2.0.13__py3-none-any.whl

agent_server/langchain/custom_middleware.py

@@ -11,10 +11,13 @@ import re
 import uuid
 from typing import Any, Dict, Optional
 
+from json_repair import repair_json
 from langchain_core.messages import AIMessage, HumanMessage
 
 from agent_server.langchain.logging_utils import (
     _format_middleware_marker,
+    _pretty_json,
+    _serialize_message,
     _with_middleware_logging,
 )
 from agent_server.langchain.prompts import JSON_TOOL_SCHEMA, NON_HITL_TOOLS
@@ -75,6 +78,15 @@ def parse_json_tool_call(text) -> Optional[Dict[str, Any]]:
         except json.JSONDecodeError:
             pass
 
+    # Try json-repair for malformed JSON from LLMs
+    try:
+        repaired = repair_json(text, return_objects=True)
+        if isinstance(repaired, dict) and "tool" in repaired:
+            logger.info(f"Repaired malformed JSON tool call: {repaired.get('tool')}")
+            return repaired
+    except Exception as e:
+        logger.debug(f"json-repair failed: {e}")
+
     return None
 
 
@@ -83,15 +95,110 @@ def normalize_tool_name(tool_name: str) -> str:
 
     Rules:
     - write_todos_tool → write_todos (TodoListMiddleware exception)
+    - task → task_tool (SubAgentMiddleware uses task_tool)
     - other tools without _tool suffix → add _tool suffix
     """
     if tool_name == "write_todos_tool":
         return "write_todos"
-    if not tool_name.endswith("_tool") and tool_name != "write_todos":
+    if tool_name == "task":
+        return "task_tool"
+    if not tool_name.endswith("_tool") and tool_name not in ("write_todos",):
         return f"{tool_name}_tool"
     return tool_name
 
 
+def repair_tool_call_arguments(arguments: str) -> Optional[Dict[str, Any]]:
+    """Repair malformed tool call arguments from LLMs.
+
+    Some LLMs (e.g., gpt-oss-120b) return arguments without leading '{' or
+    with other JSON formatting issues.
+
+    Args:
+        arguments: Raw arguments string from LLM
+
+    Returns:
+        Parsed dictionary or None if repair failed
+    """
+    if not arguments or not isinstance(arguments, str):
+        return None
+
+    arguments = arguments.strip()
+
+    # Try direct parse first
+    try:
+        return json.loads(arguments)
+    except json.JSONDecodeError:
+        pass
+
+    # Use json-repair for malformed arguments
+    try:
+        repaired = repair_json(arguments, return_objects=True)
+        if isinstance(repaired, dict):
+            logger.info("Repaired malformed tool arguments")
+            return repaired
+    except Exception as e:
+        logger.debug(f"json-repair failed for arguments: {e}")
+
+    return None
+
+
+def try_extract_tool_calls_from_additional_kwargs(
+    response_message,
+) -> Optional[AIMessage]:
+    """Try to extract and repair tool_calls from additional_kwargs.
+
+    Some LLMs put tool_calls in additional_kwargs but with malformed arguments.
+    This function tries to repair them and create a proper AIMessage.
+
+    Args:
+        response_message: AIMessage with potential tool_calls in additional_kwargs
+
+    Returns:
+        New AIMessage with repaired tool_calls, or None if extraction failed
+    """
+    if not response_message:
+        return None
+
+    additional_kwargs = getattr(response_message, "additional_kwargs", {})
+    raw_tool_calls = additional_kwargs.get("tool_calls", [])
+
+    if not raw_tool_calls:
+        return None
+
+    repaired_tool_calls = []
+    for tc in raw_tool_calls:
+        func = tc.get("function", {})
+        name = func.get("name")
+        arguments = func.get("arguments", "")
+        tc_id = tc.get("id", str(uuid.uuid4()))
+
+        if not name:
+            continue
+
+        # Try to repair arguments
+        args = repair_tool_call_arguments(arguments)
+        if args is not None:
+            repaired_tool_calls.append(
+                {
+                    "name": normalize_tool_name(name),
+                    "args": args,
+                    "id": tc_id,
+                    "type": "tool_call",
+                }
+            )
+
+    if repaired_tool_calls:
+        logger.info(
+            f"Extracted {len(repaired_tool_calls)} tool calls from additional_kwargs"
+        )
+        return AIMessage(
+            content=getattr(response_message, "content", "") or "",
+            tool_calls=repaired_tool_calls,
+        )
+
+    return None
+
+
 def create_tool_call_message(tool_name: str, arguments: Dict[str, Any]) -> AIMessage:
     """Create AIMessage with tool_calls from parsed JSON.
 
@@ -125,7 +232,7 @@ def create_handle_empty_response_middleware(wrap_model_call):
     1. Detects empty or text-only responses (no tool_calls)
     2. Retries with JSON schema prompt to force structured output
     3. Parses JSON response and injects tool_calls into AIMessage
-    4. Falls back to synthetic final_answer if all else fails
+    4. Falls back to synthetic write_todos completion if all else fails
 
     Args:
         wrap_model_call: LangChain's wrap_model_call decorator
@@ -142,65 +249,129 @@ def create_handle_empty_response_middleware(wrap_model_call):
         # Check if all todos are completed - if so, return empty response to stop agent
         # Method 1: Check state.todos
         todos = request.state.get("todos", [])
+        logger.debug(
+            "handle_empty_response: state.todos=%s",
+            json.dumps(todos, ensure_ascii=False) if todos else "[]",
+        )
         if todos:
             pending_todos = [
                 t for t in todos if t.get("status") in ("pending", "in_progress")
             ]
             if not pending_todos:
-                logger.info(
-                    "All %d todos completed (from state) - stopping agent (no LLM call)",
-                    len(todos),
+                # Check if summary already exists AFTER the last REAL HumanMessage
+                # (to avoid false positives from previous task summaries)
+                # Note: Skip system-injected messages like "[SYSTEM] Tool completed..."
+                summary_exists = False
+                messages = request.messages
+
+                # Find index of last REAL HumanMessage (not system-injected)
+                last_human_idx = -1
+                for i, msg in enumerate(messages):
+                    msg_type = getattr(msg, "type", "") or type(msg).__name__
+                    if msg_type in ("human", "HumanMessage"):
+                        msg_content = getattr(msg, "content", "") or ""
+                        # Skip system-injected messages
+                        if not msg_content.startswith("[SYSTEM]"):
+                            last_human_idx = i
+
+                # Only check messages after last REAL HumanMessage for summary
+                messages_to_check = (
+                    messages[last_human_idx + 1 :]
+                    if last_human_idx >= 0
+                    else messages[-10:]
                 )
-                return AIMessage(content="", tool_calls=[])
+                for msg in messages_to_check:
+                    content = getattr(msg, "content", "") or ""
+                    if '"summary"' in content and '"next_items"' in content:
+                        summary_exists = True
+                        break
+
+                if summary_exists:
+                    logger.info(
+                        "All %d todos completed and summary exists after last user message - stopping agent (no LLM call)",
+                        len(todos),
+                    )
+                    return AIMessage(content="", tool_calls=[])
+                else:
+                    # Allow one more LLM call for summary generation
+                    logger.info(
+                        "All %d todos completed but no summary yet after last user message - allowing LLM call for summary",
+                        len(todos),
+                    )
 
         # Method 2: Check last message if it's a write_todos ToolMessage with all completed
-        messages = request.messages
-        if messages:
-            last_msg = messages[-1]
-            if getattr(last_msg, "type", "") == "tool":
-                tool_name = getattr(last_msg, "name", "") or ""
-                content = getattr(last_msg, "content", "") or ""
-                if tool_name == "write_todos" or "Updated todo list to" in content:
-                    # Extract todos from ToolMessage content
-                    try:
-                        import ast
-                        if "Updated todo list to" in content:
-                            list_text = content.split("Updated todo list to", 1)[1].strip()
-                            todos_from_msg = ast.literal_eval(list_text)
-                            if isinstance(todos_from_msg, list) and len(todos_from_msg) > 0:
-                                pending = [t for t in todos_from_msg if t.get("status") in ("pending", "in_progress")]
-                                if not pending:
-                                    logger.info(
-                                        "All %d todos completed (from ToolMessage) - stopping agent (no LLM call)",
-                                        len(todos_from_msg),
-                                    )
-                                    return AIMessage(content="", tool_calls=[])
-                    except Exception as e:
-                        logger.debug("Failed to parse todos from ToolMessage: %s", e)
+        # Note: We now allow one more LLM call for summary generation when all todos are completed
+        # This check is skipped to let the agent produce a summary
+
+        # Check if summary todo is completed
+        # IMPORTANT: Only consider summary completed if it's the LAST todo item and ALL todos are done
+        # This prevents false positives when a previous summary is completed but new tasks are added
+        all_todos_completed = all(t.get("status") == "completed" for t in todos)
+        last_todo_is_summary = (
+            len(todos) > 0
+            and "작업 요약" in todos[-1].get("content", "")
+            and "다음 단계" in todos[-1].get("content", "")
+            and todos[-1].get("status") == "completed"
+        )
+        summary_todo_completed = all_todos_completed and last_todo_is_summary
+
+        if not summary_todo_completed and any(
+            t.get("status") == "completed"
+            and "작업 요약" in t.get("content", "")
+            for t in todos
+        ):
+            logger.debug(
+                "Previous summary todo completed but new tasks exist - NOT treating as final summary"
+            )
 
-        # Check if last message is final_answer_tool result - if so, don't retry/synthesize
-        # This allows agent to naturally terminate after final_answer_tool
+        # Check if summary content exists in messages
         messages = request.messages
-        if messages:
-            last_msg = messages[-1]
-            if getattr(last_msg, "type", "") == "tool":
-                tool_name = getattr(last_msg, "name", "") or ""
-                if not tool_name:
-                    try:
-                        content_json = json.loads(last_msg.content)
-                        tool_name = content_json.get("tool", "")
-                    except (json.JSONDecodeError, TypeError, AttributeError):
-                        pass
-                if tool_name in ("final_answer_tool", "final_answer"):
-                    logger.info(
-                        "Last message is final_answer_tool result - allowing natural termination"
-                    )
-                    # Just call handler and return response as-is (no retry/synthesize)
-                    return handler(request)
+        summary_exists = False
+        for msg in messages[-15:]:
+            msg_content = getattr(msg, "content", "") or ""
+            if '"summary"' in msg_content and '"next_items"' in msg_content:
+                summary_exists = True
+                break
+            if any(
+                kw in msg_content
+                for kw in [
+                    "다음 단계 제안",
+                    "다음 단계:",
+                    "### 다음 단계",
+                    "## 다음 단계",
+                    "**다음 단계**",
+                    "모든 작업이 완료",
+                    "**작업 요약**",
+                    "### 작업 요약",
+                    "## 작업 요약",
+                ]
+            ):
+                summary_exists = True
+                break
 
         for attempt in range(max_retries + 1):
             response = handler(request)
 
+            # If summary todo is completed AND summary content exists, accept empty response
+            # This prevents infinite loop when inject_continuation_middleware returns empty AIMessage
+            response_message = _extract_ai_message(response)
+            if summary_todo_completed and summary_exists:
+                has_content_check = (
+                    bool(getattr(response_message, "content", None))
+                    if response_message
+                    else False
+                )
+                has_tool_calls_check = (
+                    bool(getattr(response_message, "tool_calls", None))
+                    if response_message
+                    else False
+                )
+                if not has_content_check and not has_tool_calls_check:
+                    logger.info(
+                        "Summary todo completed AND summary exists - accepting empty response (agent should stop)"
+                    )
+                    return response
+
             # Extract AIMessage from response
             response_message = _extract_ai_message(response)
 
@@ -227,6 +398,23 @@ def create_handle_empty_response_middleware(wrap_model_call):
             if has_tool_calls:
                 return response
 
+            # Try to extract and repair tool_calls from additional_kwargs
+            # Some LLMs (e.g., gpt-oss-120b) put tool_calls in additional_kwargs
+            # but with malformed arguments (missing '{', broken JSON, etc.)
+            if response_message and not has_tool_calls:
+                repaired_message = try_extract_tool_calls_from_additional_kwargs(
+                    response_message
+                )
+                if repaired_message and repaired_message.tool_calls:
+                    logger.info(
+                        "Repaired tool_calls from additional_kwargs: %d calls",
+                        len(repaired_message.tool_calls),
+                    )
+                    response = _replace_ai_message_in_response(
+                        response, repaired_message
+                    )
+                    return response
+
             # Try to parse JSON from content
             if has_content and response_message:
                 parsed = parse_json_tool_call(response_message.content)
@@ -242,6 +430,117 @@ def create_handle_empty_response_middleware(wrap_model_call):
                     response = _replace_ai_message_in_response(response, new_message)
                     return response
 
+                # Check if content is summary JSON (for summary todo)
+                # Summary JSON has "summary" and "next_items" but no "tool"
+                # IMPORTANT: Check for summary JSON pattern FIRST, regardless of current todo
+                # This handles cases where LLM outputs summary JSON mixed with other content
+                content = response_message.content
+                if isinstance(content, list):
+                    content = " ".join(str(p) for p in content)
+
+                # Check if content contains summary JSON pattern
+                has_summary_pattern = ('"summary"' in content or "'summary'" in content) and (
+                    '"next_items"' in content or "'next_items'" in content
+                )
+
+                if has_summary_pattern:
+                    # Try to extract and repair summary JSON from mixed content
+                    try:
+                        # Try to find JSON object containing summary
+                        import re
+                        json_match = re.search(r'\{[^{}]*"summary"[^{}]*"next_items"[^{}]*\}', content, re.DOTALL)
+                        if json_match:
+                            repaired_summary = repair_json(
+                                json_match.group(), return_objects=True
+                            )
+                        else:
+                            repaired_summary = repair_json(
+                                content, return_objects=True
+                            )
+
+                        if (
+                            isinstance(repaired_summary, dict)
+                            and "summary" in repaired_summary
+                            and "next_items" in repaired_summary
+                        ):
+                            # Create new message with repaired JSON content
+                            repaired_content = json.dumps(
+                                repaired_summary, ensure_ascii=False
+                            )
+                            logger.info(
+                                "Detected and repaired summary JSON in content (pattern-based detection)"
+                            )
+                            # Create message with repaired content
+                            repaired_response_message = AIMessage(
+                                content=repaired_content,
+                                tool_calls=getattr(
+                                    response_message, "tool_calls", []
+                                )
+                                or [],
+                            )
+                            synthetic_message = _create_synthetic_completion(
+                                request,
+                                repaired_response_message,
+                                has_content=True,
+                            )
+                            response = _replace_ai_message_in_response(
+                                response, synthetic_message
+                            )
+                            return response
+                    except Exception as e:
+                        logger.debug(f"Failed to extract summary JSON from mixed content: {e}")
+
+                    # Fallback: accept as-is if repair failed but looks like summary
+                    logger.info(
+                        "Detected summary JSON pattern in content - accepting and synthesizing write_todos"
+                    )
+                    synthetic_message = _create_synthetic_completion(
+                        request, response_message, has_content=True
+                    )
+                    response = _replace_ai_message_in_response(
+                        response, synthetic_message
+                    )
+                    return response
+
+                # Legacy: Also check if current todo is a summary todo (backward compatibility)
+                todos = request.state.get("todos", [])
+                in_progress_todos = [
+                    t for t in todos if t.get("status") == "in_progress"
+                ]
+                pending_todos = [t for t in todos if t.get("status") == "pending"]
+                current_todo = (
+                    in_progress_todos[0]
+                    if in_progress_todos
+                    else pending_todos[0]
+                    if pending_todos
+                    else None
+                )
+                if current_todo:
+                    summary_keywords = [
+                        "작업 요약",
+                        "결과 요약",
+                        "분석 요약",
+                        "요약 및",
+                        "다음단계",
+                        "다음 단계",
+                        "next step",
+                    ]
+                    is_summary_todo = any(
+                        kw in current_todo.get("content", "") for kw in summary_keywords
+                    )
+                    if is_summary_todo and ('"summary"' in content or "'summary'" in content):
+                        # This is a summary todo with summary content - accept it
+                        logger.info(
+                            "Summary todo with summary content detected - accepting"
+                        )
+                        synthetic_message = _create_synthetic_completion(
+                            request, response_message, has_content=True
+                        )
+                        response = _replace_ai_message_in_response(
+                            response, synthetic_message
+                        )
+                        return response
+
             # Invalid response - retry with JSON schema prompt
             if response_message and attempt < max_retries:
                 reason = "text-only" if has_content else "empty"
@@ -256,7 +555,7 @@ def create_handle_empty_response_middleware(wrap_model_call):
                         reason,
                     )
                     # Synthesize write_todos while preserving the content (summary)
-                    synthetic_message = _create_synthetic_final_answer(
+                    synthetic_message = _create_synthetic_completion(
                         request, response_message, has_content
                     )
                     response = _replace_ai_message_in_response(
@@ -271,6 +570,8 @@ def create_handle_empty_response_middleware(wrap_model_call):
                     attempt + 1,
                     max_retries + 1,
                 )
+                if reason == "text-only":
+                    _log_invalid_ai_message(response_message, reason)
 
                 request = request.override(
                     messages=request.messages + [HumanMessage(content=json_prompt)]
@@ -293,7 +594,7 @@ def create_handle_empty_response_middleware(wrap_model_call):
                 logger.warning(
                     "Max retries exhausted. Synthesizing write_todos to complete."
                 )
-                synthetic_message = _create_synthetic_final_answer(
+                synthetic_message = _create_synthetic_completion(
                     request, response_message, has_content
                 )
                 response = _replace_ai_message_in_response(response, synthetic_message)
@@ -325,6 +626,23 @@ def _extract_ai_message(response):
     return None
 
 
+def _log_invalid_ai_message(response_message, reason: str) -> None:
+    """Log full AIMessage details for invalid (text-only) responses."""
+    if not response_message:
+        return
+    try:
+        payload = _serialize_message(response_message)
+    except Exception as exc:
+        logger.warning(
+            "Invalid AIMessage detail (%s): failed to serialize (%s). Raw=%r",
+            reason,
+            exc,
+            response_message,
+        )
+        return
+    logger.warning("Invalid AIMessage detail (%s): %s", reason, _pretty_json(payload))
+
+
 def _replace_ai_message_in_response(response, new_message):
     """Replace AIMessage in response with a new one."""
     if hasattr(response, "result"):
@@ -344,25 +662,51 @@ def _build_json_prompt(request, response_message, has_content):
     pending_todos = [t for t in todos if t.get("status") in ("pending", "in_progress")]
     in_progress_todos = [t for t in todos if t.get("status") == "in_progress"]
 
+    # Check if CURRENT todo (first in_progress or first pending) is summary/next_steps
+    # Not checking ALL pending todos - only the one we should be working on now
+    summary_keywords = [
+        "작업 요약",
+        "결과 요약",
+        "분석 요약",
+        "요약 및",
+        "다음단계",
+        "다음 단계",
+        "next step",
+    ]
+    current_todo = (
+        in_progress_todos[0]
+        if in_progress_todos
+        else pending_todos[0]
+        if pending_todos
+        else None
+    )
+    is_summary_todo = current_todo is not None and any(
+        kw in current_todo.get("content", "") for kw in summary_keywords
+    )
+
     if has_content:
         # If all todos completed, don't force another tool call
         if todos and not pending_todos:
             return None  # Signal to skip retry
 
-        # If current in_progress todo is "작업 요약 및 다음단계 제시", accept text-only response
+        # If current todo is "작업 요약 및 다음단계 제시", accept text-only response
         # The LLM is outputting the summary, we'll synthesize write_todos
-        if in_progress_todos:
-            current_todo = in_progress_todos[0].get("content", "")
-            if (
-                "작업 요약" in current_todo
-                or "다음단계" in current_todo
-                or "다음 단계" in current_todo
-            ):
-                logger.info(
-                    "Current todo is summary/next steps ('%s'), accepting text-only response",
-                    current_todo[:30],
-                )
-                return None  # Signal to skip retry - will synthesize write_todos with content
+        if is_summary_todo:
+            summary_todo = next(
+                (
+                    t
+                    for t in pending_todos
+                    if any(kw in t.get("content", "") for kw in summary_keywords)
+                ),
+                {"content": "summary"},
+            )
+            logger.info(
+                "Current todo is summary/next steps ('%s'), accepting text-only response",
+                summary_todo.get("content", "")[:30],
+            )
+            return (
+                None  # Signal to skip retry - will synthesize write_todos with content
+            )
 
         return (
             f"{JSON_TOOL_SCHEMA}\n\n"
@@ -370,6 +714,18 @@ def _build_json_prompt(request, response_message, has_content):
             f"Call the next appropriate tool to continue.\n"
             f'Example: {{"tool": "jupyter_cell_tool", "arguments": {{"code": "print(\'hello\')"}}}}'
         )
+    elif is_summary_todo:
+        # Empty response but current todo is summary - force summary JSON output
+        logger.info(
+            "Empty response but current todo is summary/next steps - forcing summary JSON prompt"
+        )
+        return (
+            f"{JSON_TOOL_SCHEMA}\n\n"
+            f"You MUST output a summary JSON with next_items. This is the final step.\n"
+            f"출력 형식 (반드시 이 형식으로 출력):\n"
+            f'{{"summary": "완료된 작업 요약 (한국어)", "next_items": [{{"subject": "다음 작업 제목", "description": "설명"}}]}}\n\n'
+            f"Do NOT call any tool. Just output the summary JSON directly in your response."
+        )
     elif pending_todos:
         todo_list = ", ".join(t.get("content", "")[:20] for t in pending_todos[:3])
         example_json = '{"tool": "jupyter_cell_tool", "arguments": {"code": "import pandas as pd\\ndf = pd.read_csv(\'titanic.csv\')\\nprint(df.head())"}}'
@@ -400,7 +756,7 @@ def _build_json_prompt(request, response_message, has_content):
         )
 
 
-def _create_synthetic_final_answer(request, response_message, has_content):
+def _create_synthetic_completion(request, response_message, has_content):
     """Create synthetic write_todos call to mark all todos as completed.
 
     This triggers automatic session termination via router's all_todos_completed check.
@@ -408,6 +764,15 @@ def _create_synthetic_final_answer(request, response_message, has_content):
     """
     todos = request.state.get("todos", [])
 
+    # Warn if there are pending todos being force-completed
+    pending_count = sum(1 for t in todos if t.get("status") == "pending")
+    if pending_count > 0:
+        logger.warning(
+            "Force-completing %d pending todos that were never started: %s",
+            pending_count,
+            [t.get("content") for t in todos if t.get("status") == "pending"]
+        )
+
     # Mark all todos as completed
     completed_todos = (
         [{**todo, "status": "completed"} for todo in todos]
@@ -614,80 +979,63 @@ def create_normalize_tool_args_middleware(wrap_model_call, tools=None):
                                             )
                                             args[key] = normalized_value
 
-                                # Ensure write_todos includes summary todo as last item
+                                # Validate write_todos: Only ONE item should be in_progress at a time
                                 if tool_name == "write_todos" and "todos" in args:
                                     todos = args["todos"]
                                     if isinstance(todos, list) and len(todos) > 0:
-                                        # Check if any todo contains summary keywords
-                                        summary_keywords = [
-                                            "작업 요약",
-                                            "다음단계",
-                                            "다음 단계",
-                                            "요약 및",
-                                        ]
-                                        has_summary = any(
-                                            any(
-                                                kw in todo.get("content", "")
-                                                for kw in summary_keywords
-                                            )
-                                            for todo in todos
-                                            if isinstance(todo, dict)
+                                        # Validate: Only ONE item should be in_progress at a time
+                                        # If multiple in_progress, keep only the first one
+                                        in_progress_count = sum(
+                                            1
+                                            for t in todos
+                                            if isinstance(t, dict)
+                                            and t.get("status") == "in_progress"
                                         )
-
-                                        if not has_summary:
-                                            # Add summary todo as last item
-                                            summary_todo = {
-                                                "content": "작업 요약 및 다음단계 제시",
-                                                "status": "pending",
-                                            }
-                                            todos.append(summary_todo)
-                                            logger.info(
-                                                "Auto-added '작업 요약 및 다음단계 제시' to write_todos (total: %d todos)",
-                                                len(todos),
-                                            )
-
-                                        # Log warning if summary todo is completed without JSON (but don't block)
-                                        for todo in todos:
-                                            if not isinstance(todo, dict):
-                                                continue
-                                            content = todo.get("content", "")
-                                            status = todo.get("status", "")
-                                            is_summary_todo = any(
-                                                kw in content for kw in summary_keywords
-                                            )
-                                            if is_summary_todo and status == "completed":
-                                                # Check if AIMessage content has summary JSON
-                                                msg_content = getattr(msg, "content", "") or ""
-                                                if isinstance(msg_content, list):
-                                                    msg_content = " ".join(
-                                                        str(p) for p in msg_content
-                                                    )
-                                                has_summary_json = (
-                                                    '"summary"' in msg_content
-                                                    and '"next_items"' in msg_content
-                                                ) or (
-                                                    "'summary'" in msg_content
-                                                    and "'next_items'" in msg_content
-                                                )
-                                                if not has_summary_json:
-                                                    # Just log warning - don't block completion to avoid infinite loop
-                                                    logger.warning(
-                                                        "Summary todo marked completed but no summary JSON in content. "
-                                                        "Allowing completion to proceed. Content: %s",
-                                                        msg_content[:200],
-                                                    )
+                                        if in_progress_count > 1:
+                                            found_first = False
+                                            for todo in todos:
+                                                if not isinstance(todo, dict):
+                                                    continue
+                                                if todo.get("status") == "in_progress":
+                                                    if found_first:
+                                                        # Reset subsequent in_progress to pending
+                                                        todo["status"] = "pending"
+                                                        logger.info(
+                                                            "Reset duplicate in_progress todo to pending: %s",
+                                                            todo.get("content", "")[
+                                                                :30
+                                                            ],
+                                                        )
+                                                    else:
+                                                        found_first = True
+
+                                        # NOTE: Previously had logic to revert summary todo to in_progress
+                                        # if no summary JSON was found. This caused infinite loops
+                                        # where LLM kept calling write_todos repeatedly.
+                                        # Now we let the natural termination logic handle this.
+                                        #
+                                        # NOTE: Also removed logic to preserve todos when LLM tries to delete them.
+                                        # The LLM should be able to modify todos freely when:
+                                        # - User rejects code approval
+                                        # - User changes their request
+                                        # - Code execution fails
+                                        # We rely on prompts to guide proper todo management.
 
         return response
 
     return normalize_tool_args
 
 
-def create_inject_continuation_middleware(wrap_model_call):
-    """Create middleware to inject continuation prompt after non-HITL tool execution.
+def create_continuation_control_middleware(wrap_model_call):
+    """Create unified middleware for continuation control.
 
-    Non-HITL tools execute immediately without user approval, which can cause
-    Gemini to produce empty responses. This middleware injects a system message
-    to remind the LLM to continue with the next action.
+    This middleware combines two functions:
+    1. BEFORE handler: Inject continuation prompt after non-HITL tool execution
+       - Checks for summary completion and stops if done
+       - Injects "[SYSTEM] Tool completed..." messages to guide LLM
+    2. AFTER handler: Prevent auto-continuation after summary JSON output
+       - Strips write_todos from responses containing summary JSON
+       - Prevents agent from auto-creating new todos after task completion
 
     Args:
         wrap_model_call: LangChain's wrap_model_call decorator
@@ -696,89 +1044,198 @@ def create_inject_continuation_middleware(wrap_model_call):
         Middleware function
     """
 
+    def _check_summary_exists(messages, last_real_human_idx: int) -> bool:
+        """Check if summary content exists in messages after last real user message."""
+        messages_to_check = (
+            messages[last_real_human_idx + 1 :]
+            if last_real_human_idx >= 0
+            else messages[-15:]
+        )
+        for msg in messages_to_check:
+            msg_content = getattr(msg, "content", "") or ""
+            # Check for summary JSON
+            if '"summary"' in msg_content and '"next_items"' in msg_content:
+                return True
+            # Check for markdown summary (common patterns)
+            if any(
+                kw in msg_content
+                for kw in [
+                    "다음 단계 제안",
+                    "다음 단계:",
+                    "### 다음 단계",
+                    "## 다음 단계",
+                    "**다음 단계**",
+                    "모든 작업이 완료",
+                    "**작업 요약**",
+                    "### 작업 요약",
+                    "## 작업 요약",
+                ]
+            ):
+                return True
+        return False
+
+    def _find_last_real_human_idx(messages) -> int:
+        """Find index of last real HumanMessage (not system-injected)."""
+        last_real_human_idx = -1
+        for i, msg in enumerate(messages):
+            msg_type = getattr(msg, "type", "") or type(msg).__name__
+            if msg_type in ("human", "HumanMessage"):
+                msg_content = getattr(msg, "content", "") or ""
+                if not msg_content.startswith("[SYSTEM]"):
+                    last_real_human_idx = i
+        return last_real_human_idx
+
     @wrap_model_call
-    @_with_middleware_logging("inject_continuation_after_non_hitl_tool")
-    def inject_continuation_after_non_hitl_tool(request, handler):
+    @_with_middleware_logging("continuation_control")
+    def continuation_control(request, handler):
         messages = request.messages
-        if not messages:
-            return handler(request)
-
-        last_msg = messages[-1]
-        if getattr(last_msg, "type", "") == "tool":
-            tool_name = getattr(last_msg, "name", "") or ""
-
-            # Try to extract tool name from content
-            if not tool_name:
-                try:
-                    content_json = json.loads(last_msg.content)
-                    tool_name = content_json.get("tool", "")
-                except (json.JSONDecodeError, TypeError, AttributeError):
-                    pass
-
-            if tool_name in NON_HITL_TOOLS:
-                # Method 1: Check state.todos
-                todos = request.state.get("todos", [])
-                pending_todos = [
-                    t for t in todos if t.get("status") in ("pending", "in_progress")
-                ]
 
-                # If all todos are completed, DON'T call LLM - return empty response to stop agent
-                if not pending_todos and todos:
+        # ===== BEFORE HANDLER: Inject continuation prompt =====
+        if messages:
+            last_msg = messages[-1]
+            if getattr(last_msg, "type", "") == "tool":
+                tool_name = getattr(last_msg, "name", "") or ""
+
+                # Try to extract tool name from content
+                if not tool_name:
+                    try:
+                        content_json = json.loads(last_msg.content)
+                        tool_name = content_json.get("tool", "")
+                    except (json.JSONDecodeError, TypeError, AttributeError):
+                        pass
+
+                if tool_name in NON_HITL_TOOLS:
+                    todos = request.state.get("todos", [])
+
+                    last_real_human_idx = _find_last_real_human_idx(messages)
+                    summary_exists = _check_summary_exists(
+                        messages, last_real_human_idx
+                    )
+
+                    # STOP if summary exists (regardless of todo status)
+                    if summary_exists:
+                        logger.info(
+                            "Summary exists after tool: %s - stopping agent (user must request next steps)",
+                            tool_name,
+                        )
+                        return AIMessage(content="", tool_calls=[])
+
+                    pending_todos = [
+                        t
+                        for t in todos
+                        if t.get("status") in ("pending", "in_progress")
+                    ]
+
+                    # If all todos completed but no summary yet, allow LLM call for summary
+                    if not pending_todos and todos:
+                        logger.info(
+                            "All %d todos completed, no summary yet after tool: %s - allowing LLM for summary",
+                            len(todos),
+                            tool_name,
+                        )
+
                     logger.info(
-                        "All todos completed (from state) after tool: %s - stopping agent (no LLM call)",
+                        "Injecting continuation prompt after non-HITL tool: %s",
                         tool_name,
                     )
-                    return AIMessage(content="", tool_calls=[])
 
-                # Method 2: Check ToolMessage content for write_todos
-                if tool_name == "write_todos" or "Updated todo list to" in (last_msg.content or ""):
-                    try:
-                        import ast
-                        content = last_msg.content or ""
-                        if "Updated todo list to" in content:
-                            list_text = content.split("Updated todo list to", 1)[1].strip()
-                            todos_from_msg = ast.literal_eval(list_text)
-                            if isinstance(todos_from_msg, list) and len(todos_from_msg) > 0:
-                                pending = [t for t in todos_from_msg if t.get("status") in ("pending", "in_progress")]
-                                if not pending:
-                                    logger.info(
-                                        "All %d todos completed (from ToolMessage) after tool: %s - stopping agent",
-                                        len(todos_from_msg),
-                                        tool_name,
-                                    )
-                                    return AIMessage(content="", tool_calls=[])
-                    except Exception as e:
-                        logger.debug("Failed to parse todos from ToolMessage: %s", e)
+                    # Skip continuation injection for write_todos
+                    # This prevents auto-continuation to next task after completing one
+                    # Agent will decide next action based on its own reasoning
+                    if tool_name == "write_todos":
+                        logger.info(
+                            "Skipping continuation prompt after write_todos - "
+                            "agent decides next action (pending: %d)",
+                            len(pending_todos) if pending_todos else 0,
+                        )
+                        # Don't inject continuation - let agent naturally continue or stop
+                    elif pending_todos:
+                        pending_list = ", ".join(
+                            t.get("content", "")[:30] for t in pending_todos[:3]
+                        )
+                        continuation = (
+                            f"Tool '{tool_name}' completed. "
+                            f"Continue with pending tasks: {pending_list}. "
+                            f"Call jupyter_cell_tool or the next appropriate tool."
+                        )
+                        new_messages = list(messages) + [
+                            HumanMessage(content=f"[SYSTEM] {continuation}")
+                        ]
+                        request = request.override(messages=new_messages)
+                    else:
+                        continuation = (
+                            f"Tool '{tool_name}' completed. "
+                            f"Create a todo list with write_todos if needed."
+                        )
+                        new_messages = list(messages) + [
+                            HumanMessage(content=f"[SYSTEM] {continuation}")
+                        ]
+                        request = request.override(messages=new_messages)
+
+        # ===== CALL HANDLER =====
+        response = handler(request)
+
+        # ===== AFTER HANDLER: Strip write_todos if summary JSON present =====
+        response_message = _extract_ai_message(response)
+        if not response_message:
+            return response
+
+        # Get content - handle both string and list formats
+        content = getattr(response_message, "content", "") or ""
+        if isinstance(content, list):
+            content = " ".join(
+                str(p) if isinstance(p, str) else p.get("text", "")
+                for p in content
+                if isinstance(p, (str, dict))
+            )
 
+        # Check if content contains summary JSON pattern
+        has_summary_json = '"summary"' in content and '"next_items"' in content
+
+        if has_summary_json:
+            tool_calls = getattr(response_message, "tool_calls", []) or []
+            write_todos_calls = [
+                tc for tc in tool_calls if tc.get("name") == "write_todos"
+            ]
+
+            if write_todos_calls:
                 logger.info(
-                    "Injecting continuation prompt after non-HITL tool: %s",
-                    tool_name,
+                    "Summary JSON 감지 - write_todos 호출 제거 (자동 계속 방지). "
+                    "제거된 write_todos 호출 수: %d",
+                    len(write_todos_calls),
                 )
 
-                if pending_todos:
-                    pending_list = ", ".join(
-                        t.get("content", "")[:30] for t in pending_todos[:3]
-                    )
-                    continuation = (
-                        f"Tool '{tool_name}' completed. "
-                        f"Continue with pending tasks: {pending_list}. "
-                        f"Call jupyter_cell_tool or the next appropriate tool."
-                    )
-                else:
-                    # No todos yet - let agent create them
-                    continuation = (
-                        f"Tool '{tool_name}' completed. "
-                        f"Create a todo list with write_todos if needed."
-                    )
-
-                new_messages = list(messages) + [
-                    HumanMessage(content=f"[SYSTEM] {continuation}")
+                filtered_tool_calls = [
+                    tc for tc in tool_calls if tc.get("name") != "write_todos"
                 ]
-                request = request.override(messages=new_messages)
 
-        return handler(request)
+                new_message = AIMessage(
+                    content=response_message.content,
+                    tool_calls=filtered_tool_calls,
+                    additional_kwargs=getattr(
+                        response_message, "additional_kwargs", {}
+                    ),
+                    response_metadata=getattr(
+                        response_message, "response_metadata", {}
+                    ),
+                )
+
+                response = _replace_ai_message_in_response(response, new_message)
+
+        return response
+
+    return continuation_control
+
+
+# Backward compatibility aliases
+def create_inject_continuation_middleware(wrap_model_call):
+    """Deprecated: Use create_continuation_control_middleware instead."""
+    return create_continuation_control_middleware(wrap_model_call)
+
 
-    return inject_continuation_after_non_hitl_tool
+def create_prevent_auto_continuation_middleware(wrap_model_call):
+    """Deprecated: Use create_continuation_control_middleware instead."""
+    return create_continuation_control_middleware(wrap_model_call)
 
 
 def create_patch_tool_calls_middleware(AgentMiddleware, ToolMessage, Overwrite):