feat(alias-ds): refactor code execution logic and plan organization

alias/pyproject.toml

@@ -42,7 +42,7 @@ dependencies = [
     "alembic>=1.16.1",
     "openpyxl>=3.1.5",
     "sentry-sdk[fastapi]===2.30.0",
-    "agentscope-runtime>=1.0.0",
+    "agentscope-runtime==1.0.4",
     "aiosqlite>=0.21.0",
     "asyncpg>=0.30.0",
     "itsdangerous>=2.2.0"

alias/src/alias/agent/agents/_alias_agent_base.py

@@ -41,6 +41,7 @@ def __init__(
             "static_control",
             "both",
         ] = "both",
+        plan_notebook: Optional[Any] = None,
     ):
         super().__init__(
             name=name,
@@ -52,6 +53,7 @@ def __init__(
             max_iters=max_iters,
             long_term_memory=long_term_memory,
             long_term_memory_mode=long_term_memory_mode,
+            plan_notebook=plan_notebook,
         )
 
         self.session_service = session_service

alias/src/alias/agent/agents/_data_science_agent.py

@@ -1,10 +1,8 @@
 # -*- coding: utf-8 -*-
 """Data Science Agent"""
 import asyncio
-import json
 import os
-from functools import partial
-from typing import List, Dict, Optional, Any, Type, cast, Literal
+from typing import Optional, Any, List, Type, Literal, cast
 import uuid
 
 from agentscope.formatter import FormatterBase
@@ -13,6 +11,8 @@
 from agentscope.model import ChatModelBase
 from agentscope.tool import ToolResponse
 from agentscope.tracing import trace_reply
+from agentscope.plan import PlanNotebook
+
 from loguru import logger
 from pydantic import BaseModel, ValidationError, Field
 from tenacity import retry, stop_after_attempt, wait_fixed
@@ -23,15 +23,17 @@
 from alias.agent.agents.common_agent_utils import (
     get_user_input_to_mem_pre_reply_hook,
 )
+
+from alias.agent.agents.ds_agent_utils.prompt_selector.llm_prompt_selector import (  # noqa: E501
+    LLMPromptSelector,
+)
+
+from alias.agent.agents.ds_agent_utils.utils import set_workspace_dir
 from .ds_agent_utils import (
     ReportGenerator,
-    LLMPromptSelector,
-    todo_write,
     get_prompt_from_file,
     files_filter_pre_reply_hook,
     add_ds_specific_tool,
-    set_run_ipython_cell,
-    install_package,
 )
 from .ds_agent_utils.ds_config import PROMPT_DS_BASE_PATH
 
@@ -68,31 +70,15 @@ def __init__(
             max_iters=max_iters,
             session_service=session_service,
             state_saving_dir=state_saving_dir,
+            plan_notebook=PlanNotebook(),
         )
 
-        install_package(self.toolkit.sandbox)
-        set_run_ipython_cell(self.toolkit.sandbox)
+        set_workspace_dir(self.toolkit.sandbox)
 
         self.uploaded_files: List[str] = []
 
-        self.todo_list: List[Dict[str, Any]] = []
-
-        self.infer_trajectories: List[List[Msg]] = []
-
-        self.detailed_report_path = os.path.join(
-            tmp_file_storage_dir,
-            "detailed_report.html",
-        )
         self.tmp_file_storage_dir = tmp_file_storage_dir
 
-        self.todo_list_prompt = get_prompt_from_file(
-            os.path.join(
-                PROMPT_DS_BASE_PATH,
-                "_agent_todo_reminder_prompt.md",
-            ),
-            False,
-        )
-
         self._sys_prompt = get_prompt_from_file(
             os.path.join(
                 PROMPT_DS_BASE_PATH,
@@ -142,17 +128,6 @@ def __init__(
         )
         self._selected_scenario_prompts: str = ""
 
-        self.toolkit.register_tool_function(
-            partial(todo_write, agent=self),
-            func_description=get_prompt_from_file(
-                os.path.join(
-                    PROMPT_DS_BASE_PATH,
-                    "_tool_todo_list_prompt.yaml",
-                ),
-                False,
-            ),
-        )
-
         self.toolkit.register_tool_function(self.think)
 
         self.register_instance_hook(
@@ -169,21 +144,14 @@ def __init__(
 
         logger.info(
             f"[{self.name}] "
-            "DeepInsightAgent initialized (fully model-driven).",
+            "DataScienceAgent initialized (fully model-driven).",
         )
 
     @property
     def sys_prompt(self) -> str:
         base_prompt = self._sys_prompt
 
-        todo_prompt = self.todo_list_prompt.replace(
-            "{todoList}",
-            json.dumps(self.todo_list, indent=2, ensure_ascii=False),
-        )
-
-        return (
-            f"{base_prompt}{self._selected_scenario_prompts}\n\n{todo_prompt}"
-        )
+        return f"{base_prompt}{self._selected_scenario_prompts}"
 
     @trace_reply
     async def reply(
@@ -429,25 +397,49 @@ async def generate_response(
             memory_log=memory_log,
         )
 
-        response, report = await report_generator.generate_report()
+        (
+            response,
+            report_md,
+            report_html,
+        ) = await report_generator.generate_report()
+        md_report_path = os.path.join(
+            self.tmp_file_storage_dir,
+            "detailed_report.md",
+        )
+        html_report_path = os.path.join(
+            self.tmp_file_storage_dir,
+            "detailed_report.html",
+        )
 
-        if report:
-            # report = report.replace(self.tmp_file_storage_dir, ".")
+        if report_html:
+            await self.toolkit.call_tool_function(
+                ToolUseBlock(
+                    type="tool_use",
+                    id=str(uuid.uuid4()),
+                    name="write_file",
+                    input={
+                        "path": md_report_path,
+                        "content": report_md,
+                    },
+                ),
+            )
             await self.toolkit.call_tool_function(
                 ToolUseBlock(
                     type="tool_use",
                     id=str(uuid.uuid4()),
                     name="write_file",
                     input={
-                        "path": self.detailed_report_path,
-                        "content": report,
+                        "path": html_report_path,
+                        "content": report_html,
                     },
                 ),
             )
             response = (
                 f"{response}\n\n"
-                "The detailed report has been saved to "
-                f"{self.detailed_report_path}."
+                "The detailed report (markdown version) has been saved to "
+                f"{md_report_path}.\n"
+                "The detailed report (html version) has been saved to "
+                f"{html_report_path}."
             )
 
         kwargs["response"] = response
@@ -518,21 +510,6 @@ async def _load_scenario_prompts(self):
         self._selected_scenario_prompts = "\n\n".join(scenario_contents)
         return self._selected_scenario_prompts
 
-    def _print_todo_list(self):
-        content = (
-            f" The todoList is :\n"
-            f"\n{json.dumps(self.todo_list, indent=4, ensure_ascii=False)}"
-            "\n" + "==" * 10 + "\n"
-        )
-        logger.log("SEND_PLAN", content)
-        with open(
-            self.session_service.log_storage_path,
-            "a",
-            encoding="utf-8",
-        ) as file:
-            # Append the content
-            file.write(content)
-
     def think(self, response: str):
         """
         Invoke this function whenever you need to
@@ -576,8 +553,8 @@ def think(self, response: str):
 def init_ds_toolkit(full_toolkit: AliasToolkit) -> AliasToolkit:
     ds_toolkit = AliasToolkit(full_toolkit.sandbox, add_all=False)
     ds_tool_list = [
+        "read_file",
         "write_file",
-        "run_ipython_cell",
         "run_shell_command",
     ]
     share_tools(full_toolkit, ds_toolkit, ds_tool_list)

alias/src/alias/agent/agents/ds_agent_utils/__init__.py

@@ -1,11 +1,9 @@
 # -*- coding: utf-8 -*-
 from .report_generation import ReportGenerator
-from .todoWrite import todo_write
 from .utils import (
     model_call_with_retry,
     set_run_ipython_cell,
     get_prompt_from_file,
-    install_package,
 )
 from .ds_toolkit import add_ds_specific_tool
 from .prompt_selector import LLMPromptSelector
@@ -15,11 +13,9 @@
 
 __all__ = [
     "ReportGenerator",
-    "todo_write",
     "model_call_with_retry",
     "get_prompt_from_file",
     "set_run_ipython_cell",
-    "install_package",
     "add_ds_specific_tool",
     "LLMPromptSelector",
     "files_filter_pre_reply_hook",

alias/src/alias/agent/agents/ds_agent_utils/built_in_prompt/_agent_system_workflow_prompt.md

@@ -7,7 +7,7 @@ You are an interactive coding assistant specialized in completing data science t
 When executing any data science task (data loading, cleaning, analysis, modeling, visualization, etc.), you **must** complete the following five steps **in order**:
 
 1. **Task Planning**
-   - Use the `todo_write` tool to break down the task and list todos.
+   - Use the `create_plan` tool to break down the task into structured subtasks.
    - Execution without planning is considered a **violation**.
 
 2. **Data Inspection**
@@ -24,16 +24,46 @@ When executing any data science task (data loading, cleaning, analysis, modeling
    - Upon successful completion or when objectively impossible to proceed (due to missing data, tool failure, etc.), call `generate_response` to formally end.
    - Do not terminate or exit silently without cause.
 
-> **Note**: `<system-reminder>` tags may appear in tool outputs or user messages, containing important system prompts. However, this content is not part of the actual user input or tool result.
-
 ---
 
 ## Task Management Rules
 
-- **You must use `todo_write` to track progress**, especially for multi-step tasks.
-- Mark each subtask as complete **immediately** upon finishing—no delays or batch updates.
+- **You must use `create_plan` to create a task plan**, especially for multi-step tasks.
+- Use `update_subtask_state` to mark subtasks as 'in_progress' when starting them.
+- Use `finish_subtask` to mark subtasks as 'done' with specific outcomes upon completion.
+- Use `revise_current_plan` to adapt the plan accordingly if new insights or results emerge.
+- Use `finish_plan` to finalize the entire task when all subtasks are complete.
 - Skipping planning risks missing critical steps—this is unacceptable.
 
+---
+## Code Execution Protocols
+
+**For all Python code execution:**
+
+1. **Write code to file first** using `write_file` or similar tools
+2. **Execute via shell** using `run_shell_command` with `python <filename>.py`
+3. **For large and key results**: Save to file (CSV/JSON/pickle), load in next step. Avoid printing large outputs or passing data via stdout
+4. **Never execute code inline or outside this workflow**
+
+**Example:**
+
+Tool: `write_file`
+Arguments: {
+"file_path": "/workspace/code/analysis.py",
+"content": "import pandas as pd\ndf = pd.read_csv('data.csv')\nprint(df.head())"
+}
+
+Tool: `run_shell_command`
+Arguments: {
+"command": "python /workspace/code/analysis.py"
+}
+
+Tool: `write_file`
+Arguments: {
+"file_path": "/workspace/code/step1_aggregate.py",
+"content": "import pandas as pd\ndf = pd.read_csv('raw_data.csv')\nprocessed = df.groupby('category').sum()\nprocessed.to_csv('/workspace/data/intermediate_result.csv', index=False)\nprint('✓ Results saved')"
+}
+
 ---
 
 ## Data Handling Requirements