# Copyright (c) 2025 Huawei Technologies Co., Ltd. All rights reserved.
"""
Planner Agent for Multi-Agent Task Coordination
This agent serves as a coordinator for complex tasks that require multiple agents
working together. It implements the ReAct pattern for reasoning and action.
"""
import time
import json
import requests
import os
from typing import Dict, Any, List
from concurrent.futures import ThreadPoolExecutor
# Base imports
from .base_agent import BaseAgent, AgentConfig, AgentResponse, WriterAgentTaskInput
# Import agent creators for built-in task assignment
from .writer_agent import create_writer_agent
class PlannerAgent(BaseAgent):
"""
PlannerAgent coordinates multiple agents to handle complex user queries.
The agent uses the ReAct pattern (Reasoning + Acting) to analyze user requests,
break them down into manageable tasks, and coordinate the appropriate agents
to complete the work.
"""
def __init__(self, config: AgentConfig = None, shared_mcp_client=None):
# Set default agent name if not specified
if config and not config.agent_name:
config.agent_name = "PlannerAgent"
elif not config:
config = AgentConfig(agent_name="PlannerAgent")
super().__init__(config, shared_mcp_client)
# Planner-specific state
self.execution_plan = []
self.task_queue = []
# Add built-in task assignment methods to available tools
self._add_builtin_assignment_tools()
# Regenerate tool schemas with built-in assignment tools
self.tool_schemas = self._build_tool_schemas()
self.sub_agent_configs = {}
def _add_builtin_assignment_tools(self):
"""Add built-in task assignment methods as available tools"""
# Add assignment methods that share the MCP client connection
self.available_tools.update({
"assign_subjective_task_to_writer": self.assign_subjective_task_to_writer, # assign_subjective_task_to_writer
"assign_multi_objective_tasks_to_info_seeker": self.assign_multi_objective_tasks_to_info_seeker,
"assign_multi_subjective_tasks_to_info_seeker": self.assign_multi_subjective_tasks_to_info_seeker
})
def assign_multi_objective_tasks_to_info_seeker(
self,
tasks: List[Dict[str, str]],
max_workers: int = 5
) -> Dict[str, Any]:
"""
Creates multiple TaskInput objects and routes them to info_seeker agents for concurrent execution.
This tool enables the PlannerAgent to assign multiple research tasks through the MCP tool interface.
Args:
tasks: List of task dictionaries with the following keys:
- task_content (required): The specific task content
- task_steps_for_reference: Optional reference steps for execution
- deliverable_contents: Format of expected deliverable
- acceptance_checking_criteria: Criteria for task completion and quality
- workspace_id: Workspace ID for stored files and memory
- current_task_status: Description of current task status
max_workers: Maximum concurrent threads (default=4)
Returns:
MCPToolResult with execution results for all tasks
"""
try:
# Validate task count (1-4 tasks)
if not (1 <= len(tasks) <= 5):
return {
"success": False,
"error": f"Invalid task count ({len(tasks)}). Must assign 1~5 tasks. Please re-plan the task execution schedule or re-decompose the task."
}
# Import here to avoid circular imports
try:
from agents import TaskInput, create_objective_information_seeker
except ImportError:
from ..agents import TaskInput, create_objective_information_seeker
results = []
import threading
lock = threading.Lock()
def process_task(task: Dict[str, str]):
"""Process a single task with thread-safe result collection"""
try:
# Create TaskInput object
task_input = TaskInput(
task_content=task["task_content"],
task_steps_for_reference=task.get("task_steps_for_reference"),
deliverable_contents=task.get("deliverable_contents"),
current_task_status=task.get("current_task_status"),
workspace_id=None, # Session/workspace is managed by the server; no need to set explicitly
acceptance_checking_criteria=task.get("acceptance_checking_criteria")
)
# Create and execute with info seeker agent - use shared MCP client for session consistency
info_seeker_config = getattr(self, 'sub_agent_configs', {}).get('information_seeker', {})
info_seeker = create_objective_information_seeker(
model=info_seeker_config.get('model', self.config.model),
max_iterations=info_seeker_config.get('max_iterations', 30),
shared_mcp_client=self.mcp_tools.client if hasattr(self.mcp_tools, 'client') else self.mcp_tools
)
self.logger.info(f"Assigning task to InformationSeekerAgent: {task['task_content'][:8000]}...")
# Execute the task
response = info_seeker.execute_task(task_input)
if response.success:
response_data = {
"task_content": task.get("task_content", "Unknown task"),
"success": True,
"data": response.result,
"agent_name": response.agent_name,
"iterations": response.iterations,
"execution_time": response.execution_time,
# "reasoning_trace": response.reasoning_trace
}
else:
response_data = {
"task_content": task.get("task_content", "Unknown task"),
"success": False,
"error": response.error,
"agent_name": response.agent_name
}
# Thread-safe result collection
with lock:
results.append(response_data)
return response_data
except Exception as e:
error_msg = f"Task processing failed: {str(e)}"
self.logger.error(error_msg)
with lock:
results.append({
"task_content": task.get("task_content", "Unknown task"),
"success": False,
"error": error_msg
})
return None
# Execute tasks in parallel with thread pool
with ThreadPoolExecutor(max_workers=max_workers) as executor:
futures = [executor.submit(process_task, task) for task in tasks]
# Wait for all tasks to complete
for future in futures:
future.result() # Raise exceptions if any
# Check overall success
all_success = all(task_result.get("success", False) for task_result in results)
return {
"success": all_success,
"data": {"tasks": results},
"error": None if all_success else "Some tasks failed",
"metadata": {
"tool_name": "assign_multi_objective_tasks_to_info_seeker",
"task_count": len(tasks),
"success_count": sum(1 for r in results if r.get("success")),
"failure_count": sum(1 for r in results if not r.get("success"))
}
}
except Exception as e:
self.logger.error(f"Multi-task assignment failed: {e}")
return {
"success": False,
"error": f"Multi-task assignment failed: {str(e)}"
}
def assign_multi_subjective_tasks_to_info_seeker(
self,
tasks: List[Dict[str, str]],
max_workers: int = 5
) -> Dict[str, Any]:
"""
Creates multiple TaskInput objects and routes them to info_seeker agents for concurrent execution.
This tool enables the PlannerAgent to assign multiple research tasks through the MCP tool interface.
Args:
tasks: List of task dictionaries with the following keys:
- task_content (required): The specific task content
- task_steps_for_reference: Optional reference steps for execution
- deliverable_contents: Format of expected deliverable
- acceptance_checking_criteria: Criteria for task completion and quality
- workspace_id: Workspace ID for stored files and memory
- current_task_status: Description of current task status
max_workers: Maximum concurrent threads (default=4)
Returns:
MCPToolResult with execution results for all tasks
"""
try:
# Validate task count (1-4 tasks)
if not (1 <= len(tasks) <= 6):
return {
"success": False,
"error": f"Invalid task count ({len(tasks)}). Must assign 1-6 tasks."
}
# Import here to avoid circular imports
try:
from agents import TaskInput, create_subjective_information_seeker
except ImportError:
from ..agents import TaskInput, create_subjective_information_seeker
results = []
import threading
lock = threading.Lock()
def process_task(task: Dict[str, str]):
"""Process a single task with thread-safe result collection"""
try:
# Create TaskInput object
task_input = TaskInput(
task_content=task["task_content"],
task_steps_for_reference=task.get("task_steps_for_reference"),
deliverable_contents=task.get("deliverable_contents"),
current_task_status=task.get("current_task_status"),
workspace_id=self.get_session_info()["session_id"], # Session/workspace is managed by the server; no need to set explicitly
acceptance_checking_criteria=task.get("acceptance_checking_criteria")
)
# Create and execute with info seeker agent - use shared MCP client for session consistency
info_seeker_config = getattr(self, 'sub_agent_configs', {}).get('information_seeker', {})
info_seeker = create_subjective_information_seeker(
model=info_seeker_config.get('model', self.config.model),
max_iterations=info_seeker_config.get('max_iterations', 30),
shared_mcp_client=self.mcp_tools.client if hasattr(self.mcp_tools, 'client') else self.mcp_tools
)
self.logger.info(f"Assigning task to InformationSeekerAgent: {task['task_content'][:8000]}...")
# Execute the task
response = info_seeker.execute_task(task_input)
if response.success:
response_data = {
"task_content": task.get("task_content", "Unknown task"),
"success": True,
"data": response.result,
"agent_name": response.agent_name,
"iterations": response.iterations,
"execution_time": response.execution_time,
# "reasoning_trace": response.reasoning_trace
}
else:
response_data = {
"task_content": task.get("task_content", "Unknown task"),
"success": False,
"error": response.error,
"agent_name": response.agent_name
}
# Thread-safe result collection
with lock:
results.append(response_data)
return response_data
except Exception as e:
error_msg = f"Task processing failed: {str(e)}"
self.logger.error(error_msg)
with lock:
results.append({
"task_content": task.get("task_content", "Unknown task"),
"success": False,
"error": error_msg
})
return None
# Execute tasks in parallel with thread pool
with ThreadPoolExecutor(max_workers=max_workers) as executor:
futures = [executor.submit(process_task, task) for task in tasks]
# Wait for all tasks to complete
for future in futures:
future.result() # Raise exceptions if any
# Check overall success
all_success = all(task_result.get("success", False) for task_result in results)
return {
"success": all_success,
"data": {"tasks": results},
"error": None if all_success else "Some tasks failed",
"metadata": {
"tool_name": "assign_multi_subjective_tasks_to_info_seeker",
"task_count": len(tasks),
"success_count": sum(1 for r in results if r.get("success")),
"failure_count": sum(1 for r in results if not r.get("success"))
}
}
except Exception as e:
self.logger.error(f"Multi-task assignment failed: {e}")
return {
"success": False,
"error": f"Multi-task assignment failed: {str(e)}"
}
def assign_subjective_task_to_writer(
self,
task_content: str,
user_query: str,
key_files: List[Dict[str, str]]
) -> Dict[str, Any]:
"""
Assign a writing or content creation task to the WriterAgent
Args:
task_content: Detailed description of the writing task to be performed
user_query: List storing previous information seeker subtask summaries intact to preserve information from each completed research task
key_files: Curated list of relevant files with file_path and desc for each file
Returns:
Dictionary with task assignment results
"""
try:
self.logger.info("Assigning task to WriterAgent")
# Create task input
task_input = WriterAgentTaskInput(
task_content=task_content,
user_query=user_query,
key_files=key_files,
workspace_id=self.get_session_info()["session_id"],
)
# Create writer agent with shared MCP client and sub-agent configuration
writer_config = getattr(self, 'sub_agent_configs', {}).get('writer', {})
writer = create_writer_agent(
shared_mcp_client=self.mcp_tools.client,
model=writer_config.get('model', self.config.model),
max_iterations=writer_config.get('max_iterations', 20),
temperature=writer_config.get('temperature', 0.3),
max_tokens=writer_config.get('max_tokens', 16384)
)
self.logger.info(f"Assigning task to WriterAgent: {task_content[:800]}...")
# Execute the task with shared connection
response = writer.execute_task(task_input)
if response.success:
return {
"success": True,
"data": response.result,
"agent_name": response.agent_name,
"iterations": response.iterations,
"execution_time": response.execution_time,
# "reasoning_trace": response.reasoning_trace
}
else:
return {
"success": False,
"error": response.error,
"agent_name": response.agent_name
}
except Exception as e:
self.logger.error(f"Failed to assign task to WriterAgent: {e}")
return {
"success": False,
"error": f"Task assignment failed: {str(e)}"
}
def _build_system_prompt(self) -> str:
"""Build the system prompt for the planner agent"""
tool_schemas_str = json.dumps(self.tool_schemas, ensure_ascii=False)
auto_system_prompt_template = """# PlannerAgent: Multi-Agent Task Coordinator
**Role:** Analyze complex queries, first distinguish query type (long-form writing type/objective question type), then create structured plans, and coordinate specialized agents to deliver comprehensive solutions—call corresponding tools based on query type, and only invoke writer for long-form writing type queries.
#### Available Sub-Agents:
- **`information_seeker`**: Research, data gathering, web search (supports single/parallel multi-task; long-form writing type uses assign_multi_subjective_tasks_to_info_seeker, other types use assign_multi_objective_tasks_to_info_seeker)
- **`writer`**: Only invoke this sub-agent when long-form writing is required.
---
## Optimized Workflow
### 1. Query Type Judgment & Analysis & Planning Phase
**Goal:** Use the `think` tool to analyze the problem and determine whether it is a simple task (refers to tasks that do not require calling the information search agent or tool) or a complex task (requires calling info seeker). If it is a complex task, it is necessary to further analyze whether it is a objective question（do not require calling the writer agent）or a long-form writing question (requires long-form expression and need to call the writer agent later).
- **Simple Tasks:** For simple tasks that do not require info seeker invocation, you can directly call the `planner_objective_task_done` tool and write the answer in `final_answer` field without creating a todo.md file.
- **Complex Tasks:**
- For objective tasks, must use `assign_multi_objective_tasks_to_info_seeker`
- For long-form writing tasks, must use `assign_multi_subjective_tasks_to_info_seeker`, and call the writer agent to integrate the collected information to generate a very long text
- **Task Decomposition Rules:**
- Construct a task tree with a tree-like structure, where the root node represents the user's input query. Each subtask is marked with its depth in the task tree, and the entire task tree is executed from shallow to deep. Tasks at the same depth in the task tree must be independent and can be executed in parallel (via `assign_multi_xxx_tasks_to_info_seeker`) without mutual dependencies.
- At the first level of the task tree, it is essential to thoroughly design subtasks that can be executed in parallel to explore various potential background information, thereby providing more specific clues for the next step of planning.
- Competitive Redundancy Mechanism:
- For key subtasks that have a significant impact on subsequent reasoning and planning, a redundancy mechanism should be established. This involves duplicating the task at the same depth level in the task tree, enabling the parallel execution of nearly identical tasks to enhance the completion rate and robustness of the task execution.
- **Task Parallel Sending Requirements:**
- When using `assign_multi_xxx_tasks_to_info_seeker`, all parallel-sent subtasks must be independent of each other; the description of each subtask must not contain any mutual references or dependency requirements for other subtasks.
- There is no sequential execution relationship among all parallel-sent subtasks.
- **Mandatory Documentation:** Create and write `todo.md` (e.g., `todo_v1.md`) with fields:
```markdown
# Task Planning Document
## task_name: [Clear identifier]
## task_desc: [Detailed requirements - focus on WHAT not HOW]
## deliverable_contents: [Exact output format specs]
## success_criteria: [Measurable 100% completion metrics]
## context: [Background, constraints, prior results]
## task_steps_for_reference: [Tree-structured preliminary execution plan, tag tasks with the depth in task tree `[DEPTH:xx]`]
```
### 2. Execution & Iteration Phase
#### A. Unified Iteration Triggers (Shared by Both Types)
- Based on upper-layer task results, refine the next layer of planning and document it in a new version of `todo.md` (e.g., `todo_v2.md`).
- If upper-layer tasks fail/encounter challenges: Invoke the `reflect` tool for introspection (no new information acquired, only saves thoughts), adjust the plan, and re-invoke the corresponding `information_seeker` method (objective: `assign_multi_objective_tasks_to_info_seeker`; long-form writing: `assign_multi_subjective_tasks_to_info_seeker`).
- If current tasks require prior round information: Clearly specify the context of each task and referenced files (e.g., `./data/agent_output_v1.json`) when calling `information_seeker`.
- Decompose and refine clues from upper-layer results, then execute verification in parallel.
#### B. Query-Type-Specific Operations
- **Objective tasks**: No additional operations (strictly no writer invocation). Continue iterating until information meets `success_criteria`.
- **Long-form writing tasks**: Add **information sufficiency check before writer invocation**:
1. Evaluate collected information from two dimensions: quantity (e.g., "Enough case studies for 3 chapters") and comprehensiveness (e.g., "Covers both positive and negative impacts of AI on education").
2. If information is insufficient: Adjust subtask directions (e.g., "Supplement AI education failure cases") and re-invoke `assign_multi_subjective_tasks_to_info_seeker` for targeted collection.
3. If information is sufficient: Invoke the writer via `assign_subjective_task_to_writer` (provide all collected materials and `todo.md` as context).
4. If the writer returns an incomplete result: Do not assist in completing it; only feed back the current completion status to the user.
### 3. Completion & Synthesis Phase
#### A. Unified Validation & Integration (Shared by Both Types)
- **Validation**: Cross-check multi-source `information_seeker` outputs for consistency (e.g., "NBS and World Bank GDP data differ by ≤1%").
- **Integration**: Combine parallel outputs into a unified deliverable (e.g., "Merge two GDP data sources into a single table" or "Integrate writer’s report with supplementary case studies").
- **Delivery**: Output language must match the user’s query language (e.g., Chinese query → Chinese deliverable).
#### B. Query-Type-Specific Task Completion (Critical)
- **Objective tasks**: Call the `planner_objective_task_done` tool **only when** all planned tasks are completed and the final deliverable (e.g., verified data, clear answers) is ready for user delivery.
- **Long-form writing tasks**: Call the `planner_subjective_task_done` tool **only when** the writer has finished executing and the final long-form content meets the `success_criteria` in `todo.md`.
---
## Critical Protocols
1. **Dependency Management:**
- Prohibit parallel dispatch for sequential dependent tasks unless using competitive redundancy mechanism
- Convert sequential chains to parallel where possible (e.g., Hypothesis_A vs Hypothesis_B testing)
2. **File Traceability:**
- All output references use relative paths (`./data/agent_output_1.json`)
- Version `todo.md` after each iteration (e.g., `todo_v2.md`)
3. **Local File Reading Recommendations:**
- For files crawled natively, it is not recommended to directly use the `file_read` tool to read the entire content (maybe too long). Instead, the `document_qa` tool should be used to extract and verify the required information.
- For task deliverables and summary documents from sub-agents, the `file_read` tool can be used to read them.
4. The final deliverable presented to the user should be consistent with the language used in the user's question.
5. **Writer invocation**: Strictly prohibit calling the writer for objective tasks; for long-form writing tasks, **never directly answer based on collected information**—must invoke the writer to generate the final long-form content.
Below, within the tags, are the descriptions of each tool and the required fields for invocation:
$tool_schemas
For each function call, return a JSON object placed within the [unused11][unused12] tags, which includes the function name and the corresponding function arguments:
[unused11][{\"name\": , \"arguments\": }][unused12]"""
writing_system_prompt_template = """### PlannerAgent: Multi-Agent Task Coordinator
**Role:** Analyze complex queries, create structured plans, and coordinate specialized agents to deliver comprehensive solutions.
#### Available Sub-Agents:
- **`information_seeker`**: Research, data gathering, web search (supports single/parallel multi-task)
- **`writer`**: Creates content (e.g., reports, analysis, etc.), and synthesizes from existing materials
---
### Optimized Workflow
#### 1. Analysis & Planning Phase
**Goal:** Analyze the problem and determine whether it is a simple task or a complex task. If it is a complex task, it is necessary to further analyze whether it is a subject-driven question or an objective-driven question, so as to decompose the problem into multiple clear and executable subtasks according to the specific problem type. The main characteristic of objective-driven questions is that their answers are clear and verifiable entities, otherwise they are subject-driven questions.
- **Simple Tasks:** For simple tasks that do not require sub-agent invocation, you can directly answer without creating a todo.md file
- **Complex Tasks:**
- For Objective-driven tasks, Adopt *diverge-converge* strategy:
1. Use `assign_multi_subjective_tasks_to_info_seeker` call for divergent background research
2. Converge findings to define specific sub-problems
- For Subject-driven tasks, Adopt *multi-perspective* strategy:
1. Use assign_multi_subjective_tasks_to_info_seeker call for divergent multi-source exploration (each task targets independent dimensions)
2. Converge findings to define focused sub-problems addressing distinct knowledge gaps
3. When the information seeker collects information, start to call the writer agent to integrate the collected information to generate a very long text
- **Task Decomposition Rules:**
- Construct a task tree with a tree-like structure, where the root node represents the user's input query. Each subtask is marked with its depth in the task tree, and the entire task tree is executed from shallow to deep. Tasks at the same depth in the task tree must be independent and can be executed in parallel (via `assign_multi_subjective_tasks_to_info_seeker`) without mutual dependencies.
- At the first level of the task tree, it is essential to thoroughly design subtasks that can be executed in parallel to explore various potential background information, thereby providing more specific clues for the next step of planning.
- Competitive Redundancy Mechanism:
- For key subtasks that have a significant impact on subsequent reasoning and planning, a redundancy mechanism should be established. This involves duplicating the task at the same depth level in the task tree, enabling the parallel execution of nearly identical tasks to enhance the completion rate and robustness of the task execution.
- **Task Parallel Sending Requirements:**
- When using `assign_multi_subjective_tasks_to_info_seeker`, all parallel-sent subtasks must be independent of each other; the description of each subtask must not contain any mutual references or dependency requirements for other subtasks.
- There is no sequential execution relationship among all parallel-sent subtasks.
- **Mandatory Documentation:** Create and write `todo.md` (e.g., `todo_v1.md`) with fields:
```markdown
# Task Planning Document
## task_name: [Clear identifier]
## task_desc: [Detailed requirements - focus on WHAT not HOW]
## deliverable_contents: [Exact output format specs]
## success_criteria: [Measurable 100% completion metrics]
## context: [Background, constraints, prior results]
## task_steps_for_reference: [Tree-structured preliminary execution plan, tag tasks with the depth in task tree `[DEPTH:xx]`]
```
#### 2. Execution & Iteration Phase
- **Iteration Triggers:**
- Based on the execution results of the upper layer of the task tree, specify and refine the next layer and subsequent task planning, and document them in a new `todo.md` file (e.g., `todo_v2.md`).
- If there are tasks in the previous layer that have failed or encountered challenges, it is necessary to invoke `reflect` for introspection, consider more possibilities, and make new task planning and invoke `assign_multi_subjective_tasks_to_info_seeker` again.
- If the tasks sent in the current round require reference to task information from previous rounds, it is essential to clearly specify the context of each task and the files that may need to be used or referenced when calling `assign_multi_subjective_tasks_to_info_seeker`.
- For the multiple clues of the execution results from the previous layer, they should be decomposed and refined, and executed in parallel for verification.
- **Information check required before calling writer:**
- Before invoking writer, analyze collected information for sufficiency: evaluate both quantity and comprehensiveness to ensure adequate material for long article generation
- If information is insufficient, adjust subtask direction and initiate additional targeted information collection
- **When information is sufficient, invoke writer agent** via `assign_subjective_task_to_writer`
#### 3. Completion & Synthesis Phase
- **Validation:** Cross-check multi-source outputs for consistency, and Check whether the information source is sufficient
- **Integration:** Combine parallel outputs into unified deliverable
- **Delivery:** Output language must match user's query language
- When the writer agent is finished executing, planner_subjective_task_done tool needs to be called to end the current task
---
### Critical Protocols
1. **Dependency Management:**
- Prohibit parallel dispatch for sequential dependent tasks unless using competitive redundancy mechanism
- Convert sequential chains to parallel where possible (e.g., Hypothesis_A vs Hypothesis_B testing)
2. **File Traceability:**
- All output references use relative paths (`./data/agent_output_1.json`)
- Version `todo.md` after each iteration (e.g., `todo_v2.md`)
3. **Iteration Discipline:**
- Minimum 2 parallel agents for critical hypothesis-validation tasks
- Terminate only when ALL success criteria are met at 100%
5. **Usage of Think Tool:**
- `think` is a systematic tool. After receiving the response from the complex tool or before invoking any other tools, you must **first invoke the `think` tool**: to deeply reflect on the results of previous tool invocations (if any), and to thoroughly consider and plan the user's task. The `think` tool does not acquire new information; it only saves your thoughts into memory.
6. **Usage of Reflect Tool:**
`reflect` is a systematic tool. When encountering a failure in tool execution, it is necessary to invoke the reflect tool to conduct a review and revise the task plan. It does not acquire new information; it only saves your thoughts into memory.
7. Always prioritize complete solutions over partial delivery. Use parallel redundancy for critical path tasks, and convert agent disagreements into new parallel investigation branches.
8. **CRITICAL:** When you determine that the information_seeker has gathered sufficient information, you must invoke the writer agent to draft the final article in response to the user's query. You are not allowed to reply directly based on the collected information!
9.Also note that when the writing agent returns a result that shows it is not completed, you do not need to help it complete it further. You only need to feedback the current completion status to the user.
Below, within the tags, are the descriptions of each tool and the required fields for invocation:
$tool_schemas
For each function call, return a JSON object placed within the [unused11][unused12] tags, which includes the function name and the corresponding function arguments:
[unused11][{\"name\": , \"arguments\": }][unused12]"""
qa_system_prompt_template = """### PlannerAgent: Multi-Agent Task Coordinator
**Role:** Analyze complex queries, create structured plans, and coordinate specialized agents to deliver comprehensive solutions.
#### Available Sub-Agents:
- **`information_seeker`**: Research, data gathering, web search (supports single/parallel multi-task)
---
### Optimized Workflow
#### 1. Analysis & Planning Phase
**Goal:** Decompose problems into executable units with clear dependencies
- **Simple Tasks:** For simple tasks that do not require sub-agent invocation, you can directly answer and call `planner_objective_task_done` without creating a todo.md file
- **Complex Tasks:**
- **Task Decomposition Rules:**
- Construct a task tree with a tree-like structure, where the root node represents the user\'s input query. Each subtask is marked with its depth in the task tree, and the entire task tree is executed from shallow to deep. Tasks at the same depth in the task tree must be independent and can be executed in parallel (via `assign_multi_objective_tasks_to_info_seeker`) without mutual dependencies.
- At the first level of the task tree, it is essential to thoroughly design subtasks that can be executed in parallel to explore various potential background information, thereby providing more specific clues for the next step of planning.
- Competitive Redundancy Mechanism:
- For key subtasks that have a significant impact on subsequent reasoning and planning, a redundancy mechanism should be established. This involves duplicating the task at the same depth level in the task tree, enabling the parallel execution of nearly identical tasks to enhance the completion rate and robustness of the task execution.
- **Task Parallel Sending Requirements:**
- When using `assign_multi_objective_tasks_to_info_seeker`, all parallel-sent subtasks must be independent of each other; the description of each subtask must not contain any mutual references or dependency requirements for other subtasks.
- There is no sequential execution relationship among all parallel-sent subtasks.
- **Mandatory Documentation:** Create and write `todo.md` (e.g., `todo_v1.md`) with fields:
```markdown
# Task Planning Document
## task_name: [Clear identifier]
## task_desc: [Detailed requirements - focus on WHAT not HOW]
## deliverable_contents: [Exact output format specs]
## success_criteria: [Measurable 100% completion metrics]
## context: [Background, constraints, prior results]
## task_steps_for_reference: [Tree-structured preliminary execution plan, tag tasks with the depth in task tree `[DEPTH:xx]`]
```
#### 2. Execution & Iteration Phase
- **Iteration Triggers:**
- Based on the execution results of the upper layer of the task tree, specify and refine the next layer and subsequent task planning, and document them in a new `todo.md` file (e.g., `todo_v2.md`).
- If there are tasks in the previous layer that have failed or encountered challenges, it is necessary to invoke `reflect` for introspection, consider more possibilities, and make new task planning and invoke `assign_multi_objective_tasks_to_info_seeker` again.
- If the tasks sent in the current round require reference to task information from previous rounds, it is essential to clearly specify the context of each task and the files that may need to be used or referenced when calling `assign_multi_objective_tasks_to_info_seeker`.
- For the multiple clues of the execution results from the previous layer, they should be decomposed and refined, and executed in parallel for verification.
#### 3. Completion & Synthesis Phase
- **Validation:** Cross-check multi-source outputs for consistency
- **Integration:** Combine parallel outputs into unified deliverable
- **Delivery:** Output language must match user\'s query language
- **Task Completed:** The `planner_objective_task_done` can only be called when all planned tasks have been completed and the final results are ready to be delivered to the user.
---
### Critical Protocols
1. **Dependency Management:**
- Prohibit parallel dispatch for sequential dependent tasks unless using competitive redundancy mechanism
- Convert sequential chains to parallel where possible (e.g., Hypothesis_A vs Hypothesis_B testing)
2. **File Traceability:**
- All output references use relative paths (`./data/agent_output_1.json`)
- Version `todo.md` after each iteration (e.g., `todo_v2.md`)
3. **Local File Reading Recommendations:**
- For files crawled natively, it is not recommended to directly use the `file_read` tool to read the entire content (maybe too long). Instead, the `document_qa` tool should be used to extract and verify the required information.
- For task deliverables and summary documents from sub-agents, the `file_read` tool can be used to read them.
4. The final deliverable presented to the user should be consistent with the language used in the user\'s question.
Below, within the tags, are the descriptions of each tool and the required fields for invocation:
$tool_schemas
For each function call, return a JSON object placed within the [unused11][unused12] tags, which includes the function name and the corresponding function arguments:
[unused11][{\"name\": , \"arguments\": }][unused12]"""
planner_mode_system_prompt_map = {
"auto": auto_system_prompt_template,
"writing": writing_system_prompt_template,
"qa": qa_system_prompt_template
}
system_prompt = planner_mode_system_prompt_map[self.config.planner_mode].replace("$tool_schemas", tool_schemas_str)
return system_prompt
def _build_agent_specific_tool_schemas(self) -> List[Dict[str, Any]]:
"""
Build tool schemas for PlannerAgent using proper MCP architecture.
Schemas come from MCP server via client, not direct imports.
"""
# Get MCP tool schemas from server via client (proper MCP architecture)
schemas = super()._build_agent_specific_tool_schemas()
# Add schemas for built-in task assignment tools
planner_mode_builtin_tools_map = {
"auto": ["think", "reflect", "assign_multi_subjective_tasks_to_info_seeker", "assign_multi_objective_tasks_to_info_seeker", "assign_subjective_task_to_writer", "writer_subjective_task_done", "planner_subjective_task_done", "planner_objective_task_done"],
"writing": ["think", "reflect", "assign_multi_subjective_tasks_to_info_seeker", "assign_subjective_task_to_writer", "writer_subjective_task_done", "planner_subjective_task_done"],
"qa": ["think", "reflect", "assign_multi_objective_tasks_to_info_seeker", "planner_objective_task_done"],
}
builtin_assignment_schemas = [
{
"type": "function",
"function": {
"name": "think",
"description": "Use the tool to think about something. It will not obtain new information or make any changes to the repository, but just log the thought. Use it when complex reasoning or brainstorming is needed.",
"parameters": {
"type": "object",
"properties": {
"thought": {
"type": "string",
"description": "Your thoughts."
}
},
"required": ["thought"]
}
}
},
{
"type": "function",
"function": {
"name": "reflect",
"description": "When multiple attempts yield no progress, use this tool to reflect on previous reasoning and planning, considering possible overlooked clues and exploring more possibilities. It will not obtain new information or make any changes to the repository.",
"parameters": {
"type": "object",
"properties": {
"reflect": {
"type": "string",
"description": "The specific content of your reflection"
}
},
"required": ["reflect"]
}
}
},
{
"type": "function",
"function": {
"name": "assign_multi_subjective_tasks_to_info_seeker",
"description": "Assign 1~6 research or information gathering tasks to different InformationSeekerAgents for parallel execution, each task descriptions must be semantically complete and clearly provide contextual information and potentially important reference documents.",
"parameters": {
"type": "object",
"properties": {
"tasks": {
"type": "array",
"description": "List of tasks to be assigned to multiple InformationSeekerAgents",
"items": {
"type": "object",
"properties": {
"task_content": {
"type": "string",
"description": "Detailed description of the task to be performed"
},
"task_steps_for_reference": {
"type": "string",
"description": "Optional reference steps for task execution"
},
"deliverable_contents": {
"type": "string",
"description": "Expected format and content of deliverables"
},
"current_task_status": {
"type": "string",
"description": "Current status and context of the task, important documents that may be used and referenced"
},
"acceptance_checking_criteria": {
"type": "string",
"description": "Criteria for determining task completion and quality"
},
},
"required": ["task_content"]
}
}
},
"required": ["tasks"]
}
}
},
{
"type": "function",
"function": {
"name": "assign_multi_objective_tasks_to_info_seeker",
"description": "Assign 1~5 research or information gathering tasks to different InformationSeekerAgents for parallel execution, each task descriptions must be semantically complete and clearly provide contextual information and potentially important reference documents.",
"parameters": {
"type": "object",
"properties": {
"tasks": {
"type": "array",
"description": "List of tasks to be assigned to multiple InformationSeekerAgents",
"items": {
"type": "object",
"properties": {
"task_content": {
"type": "string",
"description": "Detailed description of the task to be performed, the task description must be semantically complete"
},
"task_steps_for_reference": {
"type": "string",
"description": "Optional reference steps for task execution"
},
"deliverable_contents": {
"type": "string",
"description": "Expected format and content of deliverables"
},
"current_task_status": {
"type": "string",
"description": "Current status and context of the task, important documents that may be used and referenced"
},
"acceptance_checking_criteria": {
"type": "string",
"description": "Criteria for determining task completion and quality, and the requirements in the event of task completion failure"
},
},
"required": ["task_content"]
}
}
},
"required": ["tasks"]
}
}
},
{
"type": "function",
"function": {
"name": "assign_subjective_task_to_writer",
"description": "Assign a writing or content creation task to the WriterAgent",
"parameters": {
"type": "object",
"properties": {
"user_query": {
"type": "string",
"description": "Pass in the original user question."
},
"task_content": {
"type": "string",
"description": "Integrate and synthesize provided materials to generate comprehensive long-form content exceeding 10,000 words, especially careful not to give specific details, such as an outline plan, you are only providing the writer with a general description of the task."
},
"key_files": {
"type": "array",
"items": {
"type": "object",
"properties": {
"file_path": {
"type": "string",
"description": "Relative path to the file containing research content"
}
},
"required": ["file_path"]
},
"description": "Collect all key_files returned by the information seeker for long-form content creation."
}
},
"required": ["user_query", "task_content", "key_files"]
}
}
},
{
"type": "function",
"function": {
"name": "writer_subjective_task_done",
"description": "Writer Agent task completion reporting for complete long-form content. Called after all chapters/sections are written to provide a summary of the complete long article, final completion status and analysis, and the storage path of the final consolidated article.",
"parameters": {
"type": "object",
"properties": {
"final_article_path": {
"type": "string",
"description": "The file path where the final article is saved."
},
"article_summary": {
"type": "string",
"description": "Comprehensive summary of the complete long-form article, including main themes, key points covered, and overall narrative structure.",
"format": "markdown"
},
"completion_status": {
"type": "string",
"enum": ["completed", "partial", "failed"],
"description": "Final status of the complete long-form writing task"
},
"completion_analysis": {
"type": "string",
"description": "Analysis of the overall writing project completion including: assessment of article coherence and quality, evaluation of content organization and flow, identification of any challenges in the writing process, and overall evaluation of the long-form content creation success."
}
},
"required": ["final_article_path", "article_summary", "completion_status", "completion_analysis"]
}
}
},
{
"type": "function",
"function": {
"name": "planner_subjective_task_done",
"description": "When the writer agent is executed, the task done tool is called to end the planner's task.",
"parameters": {
"type": "object",
"properties": {
"final_article_path": {
"type": "string",
"description": "The file path where the final article is saved."
},
"task_summary": {
"type": "string",
"description": "This field is mainly used to describe the main content of the article, briefly summarize it, and finally indicate the path where the final article is saved.",
"format": "markdown"
},
"task_name": {
"type": "string",
"description": "The name of the task currently assigned to the agent, usually with underscores (e.g., 'web_research_ai_trends')"
},
"completion_status": {
"type": "string",
"enum": ["completed", "partial", "failed"],
"description": "Final task status"
}
},
"required": ["final_article_path", "task_summary", "task_name", "completion_status"]
}
}
},
{
"type": "function",
"function": {
"name": "planner_objective_task_done",
"description": "Structured reporting of task completion details including summary, decisions, and final answer",
"parameters": {
"type": "object",
"properties": {
"task_summary": {
"type": "string",
"description": "Comprehensive markdown covering what the agent was asked to do, steps taken, tools used, key findings, files created, challenges",
"format": "markdown"
},
"task_name": {
"type": "string",
"description": "The name of the task currently assigned to the agent, usually with underscores (e.g., 'web_research_ai_trends')"
},
"key_files": {
"type": "array",
"items": {
"type": "object",
"properties": {
"file_path": {
"type": "string",
"description": "Relative path to created/modified file"
},
"desc": {
"type": "string",
"description": "File contents and creation purpose"
},
"is_final_output_file": {
"type": "boolean",
"description": "Whether file is primary deliverable"
}
},
"required": ["file_path", "desc", "is_final_output_file"]
},
"description": "List of key files generated or modified during the task, with their details."
},
"completion_status": {
"type": "string",
"enum": ["completed", "partial", "failed"],
"description": "Final task status"
},
"final_answer": {
"type": "string",
"description": "The final response displayed to the user",
}
},
"required": ["task_summary", "task_name", "key_files", "completion_status", "final_answer"]
}
}
},
]
used_builtin_schemas = [schema for schema in builtin_assignment_schemas if schema["function"]["name"] in planner_mode_builtin_tools_map[self.config.planner_mode]]
schemas.extend(used_builtin_schemas)
return schemas
def _execute_react_loop(self, initial_message: str, max_iterations: int = 20) -> Dict[str, Any]:
"""
Execute the ReAct loop for planning tasks
Args:
initial_message: Initial message to start the planning process
max_iterations: Maximum number of iterations to perform
Returns:
Dictionary with execution results and trace
"""
start_time = time.time()
try:
# Reset trace for new task
self.reset_trace()
# Initialize conversation history
conversation_history = []
# Build system prompt for planning
system_prompt = self._build_system_prompt()
# Add to conversation
conversation_history.append({"role": "system", "content": system_prompt})
conversation_history.append({"role": "user", "content": initial_message + " /no_think"})
iteration = 0
task_completed = False
# Get model endpoint configuration from env-backed config
from config.config import get_config
config = get_config()
model_config = config.get_custom_llm_config()
pangu_url = model_config.get('url') or os.getenv('MODEL_REQUEST_URL', '')
model_token = model_config.get('token') or os.getenv('MODEL_REQUEST_TOKEN', '')
headers = {'Content-Type': 'application/json', 'csb-token': model_token}
# ReAct Loop: Reasoning -> Acting -> Reasoning -> Acting...
while iteration < self.config.max_iterations and not task_completed:
iteration += 1
self.logger.info(f"Planning iteration {iteration}")
try:
# Get LLM response (reasoning + potential tool calls)
retry_num = 1
max_retry_num = 10
while retry_num < max_retry_num:
try:
response = requests.post(
url=pangu_url,
headers=headers,
json={
"model": self.config.model,
"chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '[unused9]系统：[unused10]' }}{% endif %}{% if message['role'] == 'system' %}{{'[unused9]系统：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'assistant' %}{{'[unused9]助手：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'tool' %}{{'[unused9]工具：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'function' %}{{'[unused9]方法：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'user' %}{{'[unused9]用户：' + message['content'] + '[unused10]'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[unused9]助手：' }}{% endif %}",
"spaces_between_special_tokens": False,
"messages": conversation_history,
"temperature": self.config.temperature,
"max_tokens": self.config.max_tokens,
},
timeout=model_config.get("timeout", 180)
)
response = response.json()
self.logger.debug(f"API response received")
break
except Exception as e:
time.sleep(3)
retry_num += 1
if retry_num == max_retry_num:
raise ValueError(str(e))
continue
assistant_message = response["choices"][0]["message"]
# Log the reasoning
try:
if assistant_message["content"]:
reasoning_content = assistant_message["content"].split("[unused16]")[-1].split("[unused17]")[0]
if len(reasoning_content) > 0:
self.log_reasoning(iteration, reasoning_content)
except Exception as e:
self.logger.warning(f"Tool call parsing error: {e}")
# Parse error, rerun
followup_prompt = f"There is a problem with the format of model generation: {e}. Please try again."
conversation_history.append({"role": "user", "content": followup_prompt + " /no_think"})
continue
def extract_tool_calls(content):
import re
if not content:
return []
tool_call_str = re.findall(r"\[unused11\]([\s\S]*?)\[unused12\]", content)
if len(tool_call_str) > 0:
try:
tool_calls = json.loads(tool_call_str[0].strip())
except:
return []
else:
return []
return tool_calls
# Add assistant message to conversation
conversation_history.append({
"role": "assistant",
"content": assistant_message["content"]
})
tool_calls = extract_tool_calls(assistant_message["content"])
# Execute tool calls if any (Acting phase)
for tool_call in tool_calls:
arguments = tool_call["arguments"]
self.logger.debug(f"Arguments is string: {isinstance(arguments, str)}")
# Check if planning is complete
if tool_call["name"] in ["planner_subjective_task_done", "planner_objective_task_done", "writer_subjective_task_done"]:
task_completed = True
self.log_action(iteration, tool_call["name"], arguments, arguments)
break
if tool_call["name"] in ["think", "reflect"]:
tool_result = {"tool_results": "You can proceed to invoke other tools if needed. "}
else:
tool_result = self.execute_tool_call(tool_call)
# Log the action using base class method
self.log_action(iteration, tool_call["name"], arguments, tool_result)
# Add tool result to conversation
conversation_history.append({
"role": "tool",
"content": json.dumps(tool_result, ensure_ascii=False, indent=2) + " /no_think"
})
# If no tool calls, encourage continued planning
if len(tool_calls) == 0:
# Add follow-up prompt to encourage action or completion
followup_prompt = (
"Continue your planning process. Use available tools to assign tasks to agents, "
"search for information, or coordinate work. When you have a complete answer, "
"call planner_subjective_task_done or planner_objective_task_done. /no_think"
)
conversation_history.append({"role": "user", "content": followup_prompt})
except Exception as e:
error_msg = f"Error in planning iteration {iteration}: {e}"
self.log_error(iteration, error_msg)
break
execution_time = time.time() - start_time
# Extract final result
if task_completed:
# Find the completion result in the trace
completion_result = None
for step in reversed(self.reasoning_trace):
if step.get("type") == "action" and step.get("tool") in ["planner_subjective_task_done",
"planner_objective_task_done"]:
completion_result = step.get("result")
break
return {
"success": True,
"data": completion_result,
"reasoning_trace": self.reasoning_trace,
"iterations": iteration,
"execution_time": execution_time
}
else:
return {
"success": False,
"error": f"Planning task not completed within {max_iterations} iterations",
"reasoning_trace": self.reasoning_trace,
"iterations": iteration,
"execution_time": execution_time
}
except Exception as e:
execution_time = time.time() - start_time if 'start_time' in locals() else 0
self.logger.error(f"Error in execute_react_loop: {e}")
return {
"success": False,
"error": str(e),
"reasoning_trace": self.reasoning_trace,
"iterations": iteration if 'iteration' in locals() else 0,
"execution_time": execution_time
}
def execute_task(self, user_query: str) -> AgentResponse:
"""
Execute a planning task for the given user query
Args:
user_query: The user's query or request
Returns:
AgentResponse with planning results and process trace
"""
start_time = time.time()
try:
self.logger.info(f"Starting planner task: {user_query}")
# Execute the planning task using ReAct pattern
result = self._execute_react_loop(
initial_message=user_query,
max_iterations=self.config.max_iterations # Reasonable limit for planning tasks
)
execution_time = time.time() - start_time
return AgentResponse(
success=result.get("success", False),
result=result.get("data"),
error=result.get("error"),
reasoning_trace=result.get("reasoning_trace", []),
iterations=result.get("iterations", 0),
execution_time=execution_time,
agent_name=self.config.agent_name
)
except Exception as e:
execution_time = time.time() - start_time
self.logger.error(f"Planner execution failed: {e}")
return AgentResponse(
success=False,
error=f"Planner execution failed: {str(e)}",
reasoning_trace=[],
iterations=0,
execution_time=execution_time,
agent_name=self.config.agent_name
)
def create_planner_agent(
model: Any = None,
sub_agent_configs: Dict[str, Dict[str, Any]] = None,
shared_mcp_client=None,
**kwargs
) -> PlannerAgent:
"""
Create a PlannerAgent instance with server-managed sessions.
Args:
model: The LLM model to use
sub_agent_configs: Configuration for sub-agents (information_seeker, writer)
shared_mcp_client: Optional shared MCP client to prevent duplicate connections
**kwargs: Additional configuration options
Returns:
Configured PlannerAgent instance
"""
# Import the enhanced config function
from .base_agent import create_agent_config
# Create agent configuration (session managed by MCP server)
config = create_agent_config(
agent_name="PlannerAgent",
model=model,
**kwargs
)
# Create planner agent with optional shared MCP client
planner = PlannerAgent(config=config, shared_mcp_client=shared_mcp_client)
# Store sub-agent configurations for use when creating sub-agents
planner.sub_agent_configs = sub_agent_configs or {
"information_seeker": {},
"writer": {}
}
return planner