hdsp-jupyter-extension 2.0.6__py3-none-any.whl → 2.0.8__py3-none-any.whl

agent_server/langchain/prompts.py

@@ -0,0 +1,150 @@
+"""
+Prompt templates for LangChain agent.
+
+Contains system prompts, JSON schema for fallback tool calling,
+and middleware-specific prompts.
+"""
+
+DEFAULT_SYSTEM_PROMPT = """You are an expert Python data scientist and Jupyter notebook assistant.
+Your role is to help users with data analysis, visualization, and Python coding tasks in Jupyter notebooks. You can use only Korean
+
+# Core Behavior
+Be concise and direct. Answer in fewer than 4 lines unless the user asks for detail.
+After working on a file, just stop - don't explain what you did unless asked.
+Avoid unnecessary introductions or conclusions.
+
+## Task Management
+Use write_todos for complex multi-step tasks (3+ steps). Mark tasks in_progress before starting, completed immediately after finishing.
+For simple 1-2 step tasks, just do them directly without todos.
+
+You MUST ALWAYS call a tool in every response. After any tool result, you MUST:
+1. Check your todo list - are there pending or in_progress items?
+2. If YES → call the next appropriate tool (jupyter_cell_tool, markdown_tool, etc.)
+3. When you suggest next steps for todo item '다음 단계 제시', you MUST create next steps in json format matching this schema:
+{
+  "next_items": [
+    {
+        "subject": "<subject for next step>",
+        "description": "<detailed description for the next step>"
+    }, ...
+  ]
+}
+4. If ALL todos are completed → call final_answer_tool with a summary
+
+## 🔴 MANDATORY: Resource Check Before Data Hanlding
+**ALWAYS call check_resource_tool FIRST** when the task involves:
+- Loading files: .csv, .parquet, .json, .xlsx, .pickle, .h5, .feather
+- Handling datasets(dataframe) with pandas, polars, dask, or similar libraries
+- Training ML models on data files
+
+## Mandatory Workflow
+1. After EVERY tool result, immediately call the next tool
+2. Continue until ALL todos show status: "completed"
+3. ONLY THEN call final_answer_tool to summarize
+4. Only use jupyter_cell_tool for Python code or when the user explicitly asks to run in a notebook cell
+5. For plots and charts, use English text only.
+
+## ❌ FORBIDDEN (will break the workflow)
+- Producing an empty response (no tool call, no content)
+- Stopping after any tool without calling the next tool
+- Ending without calling final_answer_tool
+- Leaving todos in "in_progress" or "pending" state without continuing
+
+## 📖 File Reading Best Practices
+**CRITICAL**: When exploring codebases or reading files, use pagination to prevent context overflow.
+
+**Pattern for codebase exploration:**
+1. First scan: `read_file_tool(path, limit=100)` - See file structure and key sections
+2. Targeted read: `read_file_tool(path, offset=100, limit=200)` - Read specific sections if needed
+3. Full read: Only read without limit when necessary for immediate editing
+
+**When to paginate (use offset/limit):**
+- Reading any file >500 lines
+- Exploring unfamiliar codebases (always start with limit=100)
+- Reading multiple files in sequence
+- Any research or investigation task
+
+**When full read is OK:**
+- Small files (<500 lines)
+- Files you need to edit immediately after reading
+- After confirming file size with first scan
+
+## 🔧 Code Development
+For code generation/refactoring, use LSP tools (diagnostics_tool, references_tool) to check errors and find symbol usages. Use multiedit_file_tool for multiple changes in one file.
+"""
+
+JSON_TOOL_SCHEMA = """You MUST respond with ONLY valid JSON matching this schema:
+{
+  "tool": "<tool_name>",
+  "arguments": {"arg1": "value1", ...}
+}
+
+Available tools:
+- jupyter_cell_tool: Execute Python code. Arguments: {"code": "<python_code>"}
+- markdown_tool: Add markdown cell. Arguments: {"content": "<markdown>"}
+- final_answer_tool: Complete task. Arguments: {"answer": "<summary>"}
+- write_todos: Update task list. Arguments: {"todos": [{"content": "...", "status": "pending|in_progress|completed"}]}
+- read_file_tool: Read file with pagination. Arguments: {"path": "<file_path>", "offset": 0, "limit": 500}
+- write_file_tool: Write file. Arguments: {"path": "<path>", "content": "<content>", "overwrite": false}
+- list_files_tool: List directory. Arguments: {"path": ".", "recursive": false}
+- search_workspace_tool: Search files. Arguments: {"pattern": "<regex>", "file_types": ["py"], "path": "."}
+- search_notebook_cells_tool: Search notebook cells. Arguments: {"pattern": "<regex>"}
+- execute_command_tool: Execute shell command. Arguments: {"command": "<command>", "stdin": "<input_for_prompts>"}
+- check_resource_tool: Check resources before data processing. Arguments: {"files": ["<path>"], "dataframes": ["<var_name>"]}
+
+Output ONLY the JSON object, no markdown, no explanation."""
+
+TODO_LIST_SYSTEM_PROMPT = """
+## CRITICAL WORKFLOW RULES - MUST FOLLOW:
+- NEVER stop after calling write_todos - ALWAYS make another tool call immediately
+- For simple 1-2 step tasks, just do them directly without todos.
+
+## 🔴 NEW USER MESSAGE = FRESH START:
+- When user sends a NEW message, treat it as a COMPLETELY NEW TASK
+- IGNORE any previous todo completion history - start fresh
+- Do NOT assume any work was already done based on past conversations
+- Create a NEW todo list for the new request, even if similar items existed before
+- "다음 단계 제시" from a previous task is NOT completed for the new task
+
+## Todo List Management:
+- Before complex tasks, use write_todos to create a task list
+- Update todos as you complete each step (mark 'in_progress' → 'completed')
+- Each todo item should be specific and descriptive
+- All todo items must be written in Korean
+- ALWAYS include "다음 단계 제시" as the LAST item
+
+## Task Completion Flow:
+1. When current task is done → mark it 'completed' with write_todos
+2. For "다음 단계 제시" → mark completed, then call final_answer_tool with suggestions
+
+## FORBIDDEN PATTERNS:
+❌ Calling write_todos and then stopping
+❌ Updating todo status without doing the actual work
+❌ Ending turn without calling final_answer_tool when all tasks are done
+❌ Marking a todo as 'completed' without actually executing it in THIS conversation
+"""
+
+TODO_LIST_TOOL_DESCRIPTION = """Update the task list for tracking progress.
+⚠️ CRITICAL: This tool is ONLY for tracking - it does NOT do any actual work.
+After calling this tool, you MUST IMMEDIATELY call another tool (jupyter_cell_tool, markdown_tool, or final_answer_tool).
+NEVER end your response after calling write_todos - always continue with the next action tool."""
+
+# Non-HITL tools that execute immediately without user approval
+NON_HITL_TOOLS = {
+    "markdown_tool",
+    "markdown",
+    "read_file_tool",
+    "read_file",
+    "list_files_tool",
+    "list_files",
+    "search_workspace_tool",
+    "search_workspace",
+    "search_notebook_cells_tool",
+    "search_notebook_cells",
+    "write_todos",
+    # LSP tools (read-only)
+    "diagnostics_tool",
+    "diagnostics",
+    "references_tool",
+    "references",
+}

agent_server/langchain/state.py

@@ -6,7 +6,7 @@ This state is passed through the agent execution and updated by middleware.
 """
 
 from dataclasses import dataclass
-from typing import Annotated, Any, Dict, List, Optional, TypedDict
+from typing import Annotated, Any, Dict, List, Optional, TypedDict, Union
 
 from langchain_core.messages import BaseMessage
 from langgraph.graph.message import add_messages
@@ -14,6 +14,7 @@ from langgraph.graph.message import add_messages
 
 class NotebookContext(TypedDict, total=False):
     """Current notebook context"""
+
     notebook_path: str
     cell_count: int
     imported_libraries: List[str]
@@ -24,6 +25,7 @@ class NotebookContext(TypedDict, total=False):
 
 class ExecutionResult(TypedDict, total=False):
     """Result of code execution"""
+
     success: bool
     output: str
     error_type: Optional[str]
@@ -35,6 +37,7 @@ class ExecutionResult(TypedDict, total=False):
 
 class SearchResult(TypedDict, total=False):
     """Result of code search"""
+
     file_path: str
     cell_index: Optional[int]
     line_number: Optional[int]
@@ -51,6 +54,7 @@ class AgentState(TypedDict, total=False):
     - Tool executions (execution results, search results)
     - Agent decisions (current step, plan updates)
     """
+
     # Message history - uses add_messages reducer to accumulate messages
     messages: Annotated[List[BaseMessage], add_messages]
 
@@ -87,6 +91,9 @@ class AgentState(TypedDict, total=False):
     # Detected libraries for knowledge injection
     detected_libraries: List[str]
 
+    # Resource usage context (legacy, use check_resource_tool instead)
+    resource_context: Optional[Union[Dict[str, Any], str]]
+
     # Final answer
     final_answer: Optional[str]
     is_complete: bool
@@ -96,9 +103,10 @@ class AgentState(TypedDict, total=False):
 class AgentRuntime:
     """
     Runtime context passed to middleware.
-    
+
     Contains references to executors and services needed by middleware.
     """
+
     jupyter_executor: Any = None
     notebook_searcher: Any = None
     rag_manager: Any = None
@@ -107,7 +115,7 @@ class AgentRuntime:
     workspace_root: str = "."
 
     # Execution mode
-    embedded_mode: bool = True
+    embedded_mode: bool = False
 
     # Configuration
     max_retries: int = 3
@@ -123,19 +131,20 @@ def create_initial_state(
 ) -> AgentState:
     """
     Create initial agent state from user request.
-    
+
     Args:
         user_request: Natural language request from user
         notebook_context: Current notebook state
         llm_config: LLM configuration
-        
+
     Returns:
         Initialized AgentState
     """
     return AgentState(
         messages=[],
         user_request=user_request,
-        notebook_context=notebook_context or NotebookContext(
+        notebook_context=notebook_context
+        or NotebookContext(
             notebook_path="",
             cell_count=0,
             imported_libraries=[],
@@ -154,6 +163,7 @@ def create_initial_state(
         recovery_strategy=None,
         llm_config=llm_config or {},
         detected_libraries=[],
+        resource_context=None,
         final_answer=None,
         is_complete=False,
     )

agent_server/langchain/tools/__init__.py

@@ -7,13 +7,20 @@ Tools available:
 - final_answer: Complete the task
 - read_file: Read file content
 - write_file: Write file content
+- edit_file: Edit file with string replacement
 - list_files: List directory contents
 - search_workspace: Search files in workspace
 - search_notebook_cells: Search cells in notebooks
+- execute_command_tool: Run shell commands (client-executed)
+- check_resource_tool: Check resources before data processing (client-executed)
+- diagnostics_tool: Get LSP diagnostics (errors, warnings)
+- references_tool: Find symbol references via LSP
 """
 
 from agent_server.langchain.tools.file_tools import (
+    edit_file_tool,
     list_files_tool,
+    multiedit_file_tool,
     read_file_tool,
     write_file_tool,
 )
@@ -22,10 +29,16 @@ from agent_server.langchain.tools.jupyter_tools import (
     jupyter_cell_tool,
     markdown_tool,
 )
+from agent_server.langchain.tools.lsp_tools import (
+    diagnostics_tool,
+    references_tool,
+)
+from agent_server.langchain.tools.resource_tools import check_resource_tool
 from agent_server.langchain.tools.search_tools import (
     search_notebook_cells_tool,
     search_workspace_tool,
 )
+from agent_server.langchain.tools.shell_tools import execute_command_tool
 
 __all__ = [
     "jupyter_cell_tool",
@@ -33,7 +46,13 @@ __all__ = [
     "final_answer_tool",
     "read_file_tool",
     "write_file_tool",
+    "edit_file_tool",
+    "multiedit_file_tool",
     "list_files_tool",
     "search_workspace_tool",
     "search_notebook_cells_tool",
+    "execute_command_tool",
+    "check_resource_tool",
+    "diagnostics_tool",
+    "references_tool",
 ]