fix: tasks format

Author: gusye1234

README.md

@@ -31,7 +31,14 @@
 
 ## User Cases
 
-- `Find all .py file and explain them to me`
+<div align="center">
+		  <video width="80%" controls>
+            <source src="./examples/imgs/nano-manus.mp4" type="video/mp4">
+        </video>
+  <p>nano-manus demo video</p>
+</div>
+
+- `Find all .py file and explain them to me `
 - `Give me the latest weather in SF in last 7 days and save it to csv`
 
 > Welcome to give more user cases!
@@ -70,12 +77,14 @@ uv run examples/basic_planner.py
 - [x] Execute codes and commands in your computer (`@wonderwhy-er/desktop-commander`)
 - [ ] (coming soon) Read `.pdf, .doc`
 - [ ] (coming soon) browser use
+- [ ] (coming soon) multi-model router (`claude`, `qwen`, `deepseek`...)
 
 
 
 
 
 ## Known Issues
 
+- `nano-manus` is extremely unstable! My guess is `gpt-4o` is not that good at tool use.
 - `Unable to exit`: seem like some MCPs will cause the problems of unable to exit the program when all the tasks were done.
 - `nano-manus` will operate files and run command **on the current dir of your local computer**, make sure you don't run it on some important folders.

examples/imgs/nano-manus.mp4

@@ -0,0 +1,12 @@
+def read_and_print_weather(file_path):
+    try:
+        with open(file_path, 'r') as file:
+            contents = file.read()
+            print(contents)
+    except FileNotFoundError:
+        print(f"The file {file_path} does not exist.")
+    except Exception as e:
+        print(f"An error occurred: {e}")
+
+if __name__ == "__main__":
+    read_and_print_weather('sf_weather.csv')

examples/load_weather.py

@@ -0,0 +1,2 @@
+Weather Information
+High chance of rain, 90% probability of precipitation, Rainfall expected near half an inch

examples/sf_weather.csv

@@ -1,6 +1,6 @@
 [project]
 name = "nano_manus"
-version = "0.0.2"
+version = "0.0.3"
 description = "Implementing some features of Manus with MCP"
 readme = "README.md"
 authors = [{ name = "Gus", email = "[email protected]" }]
@@ -16,5 +16,8 @@ dependencies = [
 requires = ["hatchling"]
 build-backend = "hatchling.build"
 
+[tool.hatch.build]
+exclude = ["examples/*", "docs/*", ".github/*", ".git/*", "tests/*"]
+
 [dependency-groups]
 dev = ["pytest-asyncio>=0.25.3", "pytest>=8.3.5"]

pyproject.toml

@@ -26,6 +26,8 @@ def get_async_openai_client() -> AsyncOpenAI:
 
 async def llm_complete(model: str, messages: list[dict], **kwargs) -> ChatCompletion:
     client = get_async_openai_client()
+    if "temperature" not in kwargs:
+        kwargs["temperature"] = 0.1
     response = await client.chat.completions.create(
         model=model, messages=messages, **kwargs
     )

src/nano_manus/env.py

@@ -1,8 +1,8 @@
 import re
-from ..env import CONSOLE
 
-CODE_BLOCK_PATTERN = re.compile(r"```tasks(.*?)```", re.DOTALL)
-GOAL_BLOCK_PATTERN = re.compile(r"^(.*?)```tasks", re.DOTALL)
+CODE_BLOCK_PATTERN = re.compile(r"<tasks>(.*?)</tasks>", re.DOTALL)
+TASK_BLOCK_PATTERN = re.compile(r"<subtask>(.*?)</subtask>", re.DOTALL)
+GOAL_BLOCK_PATTERN = re.compile(r"^(.*?)<tasks>", re.DOTALL)
 
 
 def parse_step(step: str) -> dict:
@@ -18,8 +18,8 @@ def parse_step(step: str) -> dict:
         return goal, []
     else:
         subtasks = code_blocks[0].strip()
-        subtasks = [l.strip() for l in subtasks.split("\n") if l.startswith("- ")]
-        expressions = [parse_subtask_expression(st[2:]) for st in subtasks]
+        subtask_blocks = TASK_BLOCK_PATTERN.findall(subtasks)
+        expressions = [parse_subtask_expression(st.strip()) for st in subtask_blocks]
         expressions = [e for e in expressions if e is not None]
     return goal, expressions
 
@@ -29,7 +29,11 @@ def parse_subtask_expression(subtask: str) -> dict:
     subtask = subtask.strip()
 
     # Split by '=' to separate result name and the rest
-    result_name, expression = [x.strip() for x in subtask.split("=")]
+    parts = subtask.split("=")
+    result_name, expression = (
+        parts[0].strip(),
+        "=".join(parts[1:]).strip(),
+    )
 
     # Extract agent_id and param from subtask(agent_id, "param")
     # Remove 'subtask(' from start and ')' from end
@@ -44,3 +48,23 @@ def parse_subtask_expression(subtask: str) -> dict:
     param = param.strip("\"'")
 
     return {"result_name": result_name, "agent_id": agent_id, "param": param}
+
+
+if __name__ == "__main__":
+    print(
+        parse_step(
+            """
+### Step 3: Write a Python script to load the CSV file.
+
+Sub-goal: Use the Terminal Agent to create a Python file that loads and prints the content of the `sf_weather.csv` file.
+
+<tasks>
+<subtask>
+result_31 = subtask(agent_1, "Create a Python file named 'load_sf_weather.py' with the following content:\n\n```\nimport csv\n\nwith open('sf_weather.csv', 'r') as file:\n    reader = csv.reader(file)\n    for row in reader:\n        print(', '.join(row))\n```\n")
+</subtask>
+</tasks>
+
+Please let me know once this step is completed or if there are any issues.
+"""
+        )
+    )

src/nano_manus/planner/parser.py

@@ -11,6 +11,7 @@ class Planner:
     def __init__(
         self, max_steps: int = 30, max_tasks: int = 60, workers: list[BaseWorker] = []
     ):
+
         self.__workers: list[BaseWorker] = workers
         self.__max_steps = max_steps
         self.__max_tasks = max_tasks

src/nano_manus/planner/planner.py

@@ -21,19 +21,25 @@
 </input>
 <output id=0>
 I need to use agent_0 to perform 3+5 and 5+8, they can be executed in parallel.
-```tasks
-- result_11 = subtask(agent_0, "Perform 3 + 5")
-- result_12 = subtask(agent_0, "Perform 5 + 8")
-```
+<tasks>
+    <subtask>
+    result_11 = subtask(agent_0, "Perform 3 + 5")
+    </subtask>
+    <subtask>
+    result_12 = subtask(agent_0, "Perform 5 + 8")
+    </subtask>
+</tasks>
 <input id=1>
 - result_11 = 8
 - result_12 = 13
 </input>
 <output id=1>
 I need to use agent_1 to perform the subtraction between result_11 and result_12
-```tasks
-- result_21 = subtask(agent_1, "Perform result_11 - result_12")
-```
+<tasks>
+    <subtask>
+    result_21 = subtask(agent_1, "Perform result_11 - result_12")
+    </subtask>
+</tasks>
 </output>
 <input id=2>
 - result_21 = -5
@@ -44,9 +50,16 @@
 <explanation>
 Above is an example of how to plan the task into a list of steps. The following is the explanation step by step:
 - For each new step, you should start with your goal in this step.
-- After goal, you should wrap the tasks of this step with code block ```tasks.
-- The content of the code block is lines of subtasks:
-    - `- RESULT_ID = subtask(AGENT_ID, TASK_DESCRIPTION)` is a notation of a subtask, where you use the agent with id AGENT_ID to perform the task TASK_DESCRIPTION, and denote the result of this subtask as RESULT_ID.
+- After goal, you should wrap the tasks of this step with tag <tasks>.
+- The content of the <tasks> is XML of subtasks:
+```
+<subtask>
+RESULT_ID = subtask(AGENT_ID, TASK_DESCRIPTION)
+</subtask>
+<subtask>
+...
+```
+- `RESULT_ID = subtask(AGENT_ID, TASK_DESCRIPTION)` is a notation of a subtask, where you use the agent with id AGENT_ID to perform the task TASK_DESCRIPTION, and denote the result of this subtask as RESULT_ID.
 - After your plan, the user will give you the result of tasks, with the format:
 ```
 - RESULT_ID1 = AGENT_RETURN_1
@@ -55,7 +68,7 @@
 ```
 - You need to decide which step to go next based on the result.
 - Remember to use the same RESULT_ID in your later steps instead the actual value.
-- When you have finished the user task or you have no more steps to go, output the final result to response and no more ```tasks is needed.
+- When you have finished the user task or you have no more steps to go, output the final result to response and no more <tasks> is needed.
 </explanation>
 </example>
 
@@ -68,7 +81,8 @@
 ## Notes
 - If you find the previous results is wrong or not useful, you can re-plan the task with same result_id.
 - You should decide if you have enough agents to perform the task, if not, you should ask the user to add more agents.
-- Make sure you give enough context to the agents in TASK_DESCRIPTION, you ```tasks block is the only way to communicate with other agents.
+- Make sure you give enough context to the agents in TASK_DESCRIPTION, you <tasks> block is the only way to communicate with other agents.
+- If you want to agent to use the previous results, make sure you told the agent about the RESULT_ID you need them to use in TASK_DESCRIPTION
 - If you have finished the task, you need to answer the user's question and restate some result if needed.
 
 Now, understand the task in user input, and plan the task into a list of steps based on the above instructions:

src/nano_manus/planner/prompts.py

@@ -8,6 +8,10 @@
 # Tools
 {tools}
 
+# Additional Context
+Below is the additional context you may need to understand the user's instruction.
+{additional_context}
+
 # Planning step by step 
 Before you start to call functions, you should always plan step by step and explain your plan in a concise way.
 
@@ -44,6 +48,9 @@ async def handle(self, instruction: str, global_ctx: dict = {}) -> str:
             CONSOLE.print(
                 f"🤖 {self.name}: I will use previouse results", additional_context
             )
+        additional_context_string = "\n".join(
+            [f"- {k}: {v}" for k, v in additional_context.items()]
+        )
         hints = [await m.hint() for m in self.__mcps]
         tool_schemas = []
         for m in self.__mcps:
@@ -54,7 +61,9 @@ async def handle(self, instruction: str, global_ctx: dict = {}) -> str:
                 find_tools[tool.name] = m.call_tool(tool.name)
 
         system_prompt = PROMPT.format(
-            system=self.overwrite_system(), tools="\n".join(hints)
+            system=self.overwrite_system(),
+            tools="\n".join(hints),
+            additional_context=additional_context_string,
         )
         messages = [
             {