name: "trae-orchestrator" description: "Orchestrates TRAE IDE for automated software development with multi-agent collaboration. Invoke when user wants to develop software using TRAE or needs automated project management."

TRAE Orchestrator

Automated software development controller that orchestrates TRAE IDE for fully autonomous project delivery.

When to Invoke

User wants to develop software using TRAE
User needs automated project management
User provides software requirements and project directory
User asks for multi-agent development workflow
User wants to automate TRAE with Python scripts

Quick Start (Recommended)

One-Line Project Launch

from automation_helper import quick_start

# 一键启动项目
quick_start(
    project_dir='D:\\MyProject',
    requirements={
        'name': '我的项目',
        'description': '项目描述...',
        'features': ['功能1', '功能2'],
        'tech_stack': 'Node.js + React'
    }
)

This will:

✅ Create project structure
✅ Create requirements.md
✅ Create prompt for TRAE
✅ Launch TRAE IDE
✅ Send development task to TRAE

Automation Helper Module

A practical Python module (automation_helper.py) is provided for easy automation:

TRAEController - IDE Controller

from automation_helper import TRAEController

# Initialize (auto-detects TRAE path)
controller = TRAEController()
# Or specify path
controller = TRAEController('E:\\software\\Trae CN\\Trae CN.exe')

# First-time setup
controller.setup('E:\\software\\Trae CN\\Trae CN.exe')

# Launch TRAE with project
controller.launch('D:\\MyProject')

# Send prompt (requires pyautogui)
controller.send_prompt("Create a web app...", delay=5)

ProjectManager - Project Setup

from automation_helper import ProjectManager

# Create project structure
ProjectManager.create_project(
    project_dir='D:\\MyProject',
    requirements={
        'name': '星空篝火游戏',
        'description': '多人联机游戏',
        'features': ['3D场景', '多人联机', '聊天系统'],
        'tech_stack': 'Three.js + Node.js'
    }
)

# Create prompt for TRAE
ProjectManager.create_prompt('D:\\MyProject')

ProgressMonitor - Monitor Progress

from automation_helper import ProgressMonitor

# Monitor project progress
monitor = ProgressMonitor('D:\\MyProject')

# Check signals
if monitor.check_signal('project_done'):
    print("Project complete!")

# Get status summary
status = monitor.get_status()
print(status)

# Wait for completion
monitor.wait_for_completion(timeout=3600)  # 1 hour timeout

User Control Functions

from automation_helper import pause_project, resume_project, stop_project

pause_project('D:\\MyProject')   # Pause
resume_project('D:\\MyProject')  # Resume
stop_project('D:\\MyProject')    # Stop

Token Optimization Strategy

CRITICAL: Minimize openclaw Token Usage

openclaw Does	TRAE Does (Free)
Orchestrate workflow	All code generation
Read only: task_plan.md, progress.md	Read/write all source files
Send prompts	Execute prompts
Detect completion	Self-check quality
Intervene on loops	Auto-fix bugs (3 attempts)

Event-Driven Completion Detection (No Polling!)

DO NOT poll every 30 seconds. Use these efficient methods:

Method 1: Signal File (Most Efficient)

TRAE creates a signal file when done - openclaw only checks if file exists:

# In prompt, instruct TRAE:
"When phase complete, create file: .trae-docs/.signal_{PHASE}_DONE"

# openclaw checks:
if os.path.exists('.trae-docs/.signal_planning_done'):
    # Phase complete, read progress.md once
    # Delete signal file after reading

Token cost: 0 (file existence check is free)

Method 2: File Modification Time

Only read when timestamp changes:

last_mtime = 0

def check_progress():
    global last_mtime
    current_mtime = os.path.getmtime('.trae-docs/progress.md')
    if current_mtime > last_mtime:
        last_mtime = current_mtime
        return read_file('.trae-docs/progress.md')
    return None  # No change, don't read

Token cost: 0 until file actually changes

Method 3: Watchdog File Monitor (Background)

Use filesystem events instead of polling:

from watchdog.observers import Observer
from watchdog.events import FileSystemEventHandler

class ProgressHandler(FileSystemEventHandler):
    def on_modified(self, event):
        if 'progress.md' in event.src_path:
            # File changed, now read it
            content = read_file(event.src_path)
            process_status(content)

observer = Observer()
observer.schedule(ProgressHandler(), path='.trae-docs/')
observer.start()

Token cost: 0 until file changes, then only 1 read

Recommended: Signal File + Timestamp Combo

┌─────────────────────────────────────────────────────────┐
│  TRAE completes task                                    │
│       ↓                                                 │
│  TRAE creates .signal_done (empty file)                 │
│       ↓                                                 │
│  openclaw detects signal file exists (0 tokens)         │
│       ↓                                                 │
│  openclaw reads progress.md once                        │
│       ↓                                                 │
│  openclaw deletes signal file                           │
│       ↓                                                 │
│  openclaw sends next prompt                             │
└─────────────────────────────────────────────────────────┘

First-Time Setup

Step 1: Get TRAE Installation Path

Ask user: "Please provide the TRAE installation directory path"
Example: "C:\Users\XXX\AppData\Local\Programs\Trae CN"

Step 2: Verify and Save

Check if directory contains Trae CN.exe
Launch TRAE to verify it works
Save to config.json:

{
  "trae_install_path": "USER_PROVIDED_PATH",
  "trae_executable": "Trae CN.exe",
  "window_identifier": "Trae CN",
  "max_instances": 3,
  "version": "1.0.0"
}

Project Structure

{project_dir}/
├── .trae-docs/
│   ├── requirements.md    # User requirements
│   ├── architecture.md    # System design
│   ├── task_plan.md       # Development plan
│   ├── progress.md        # Current status (openclaw reads this)
│   └── review_log.md      # Review history
└── src/                   # Generated code (TRAE manages)

Super-Efficient Workflow

Phase 1: Planning (One Prompt)

Send single comprehensive prompt:

Develop [SOFTWARE_TYPE] with these requirements:

[REQUIREMENTS]

Tech stack: [TECHNOLOGIES]

INSTRUCTIONS:
1. Create .trae-docs/architecture.md with system design
2. Create .trae-docs/task_plan.md with task breakdown
3. Create .trae-docs/progress.md with initial status
4. Each task must be completable within 200k tokens
5. Include acceptance criteria for each task
6. Mark task dependencies clearly

COMPLETION SIGNAL:
When done, create empty file: .trae-docs/.signal_planning_done
Also update progress.md with:
STATUS: PLANNING_COMPLETE
TASKS_TOTAL: N
ESTIMATED_TOKENS: N

Use SOLO mode. Work autonomously.

Detection: Check if .signal_planning_done exists (0 tokens), then read progress.md once.

Phase 2: Batch Implementation

Send tasks in batches (not one by one):

BATCH IMPLEMENTATION - Tasks [START_ID] to [END_ID]

Read .trae-docs/task_plan.md for task details.

For each task:
1. Implement following architecture.md
2. Write unit tests
3. Update progress.md with completion status
4. Mark task as [x] in task_plan.md

COMPLETION SIGNAL:
After ALL tasks in batch:
1. Create empty file: .trae-docs/.signal_batch_[N]_done
2. Update progress.md with:
   STATUS: BATCH_[N]_COMPLETE
   COMPLETED_TASKS: [IDs]
   REMAINING_TASKS: [IDs]

Work autonomously in SOLO mode.

Detection: Check if .signal_batch_N_done exists (0 tokens), then read progress.md once.

Phase 3: Self-Review

Let TRAE review itself:

SELF-REVIEW PHASE

Review all implemented code:
1. Check against requirements.md
2. Run all tests
3. Check code quality
4. Document issues in review_log.md

If issues found:
- Fix them automatically
- Re-run tests
- Update review_log.md

COMPLETION SIGNAL:
When done, create empty file: .trae-docs/.signal_review_done
Also update progress.md with:
STATUS: REVIEW_COMPLETE
ISSUES_FOUND: N
ISSUES_FIXED: N

If blocked, create: .trae-docs/.signal_blocked
And update progress.md with:
STATUS: BLOCKED
BLOCKER: [description]

Detection: Check if .signal_review_done or .signal_blocked exists (0 tokens), then read progress.md once.

Minimal Intervention Protocol

Intervention Triggers (Signal-Based)

Signal File	Action
`.signal_blocked`	Read blocker description, provide guidance
`.signal_need_clarification`	Ask user for input
`.signal_error_loop`	Read error log, send new approach
`.signal_context_full`	Start new conversation with checkpoint

No Intervention Needed When

No signal files present (TRAE is working)
.signal_batch_N_done exists (normal progress)
Files are being modified (active development)

Timeout Fallback

Only if no signal file and no file changes for 10+ minutes:

# Last resort check
if no_signal_files() and file_age('progress.md') > 600:
    # Check TRAE window state
    screenshot = capture_trae_window()
    if "产物汇总" in screenshot:
        # TRAE finished but forgot signal
        create_signal_file('.signal_done')
    elif is_idle(screenshot):
        # TRAE is stuck
        create_signal_file('.signal_blocked')

Error Handling

Bug-Fix Loop (3+ attempts detected via .signal_error_loop)

ALTERNATIVE APPROACH for [BUG_ID]

Previous attempts failed. Try:
1. [DIFFERENT_APPROACH]
2. Consider: [ALTERNATIVE_SOLUTION]
3. If still fails after 3 more attempts:
   - Create .signal_blocked
   - Update progress.md with BLOCKER description

Start fresh. Do not reference previous attempts.

COMPLETION SIGNAL:
- Success: Create .signal_fixed_[BUG_ID]
- Failed: Create .signal_blocked

Context Overflow (TRAE handles automatically)

Include in initial prompt:

CONTEXT MANAGEMENT:
- Monitor token usage
- When approaching 200k tokens:
  1. Create checkpoint summary in progress.md
  2. Create .signal_context_full
  3. List remaining tasks
  4. Note partial implementations

When openclaw detects .signal_context_full:

Start new TRAE conversation with:
"Continue from checkpoint. Read progress.md for context.
Remaining tasks: [LIST]
Resume from: [LAST_COMPLETED_TASK]"

Multi-Agent Strategy

When to Use Multiple TRAE Windows

Project Size	Strategy
Small (<10 tasks)	Single TRAE instance
Medium (10-30 tasks)	2 instances: Planner+Coder, Reviewer
Large (>30 tasks)	3 instances: Planner, Coder, Reviewer

Parallel Execution

For large projects, run Coder and Reviewer in parallel:

Window 1 (Coder): Implement tasks 1-5
Window 2 (Reviewer): Review completed tasks

Progress File Format

TRAE updates progress.md - openclaw only reads this file:

# Project Progress

## Status: [PLANNING|IMPLEMENTING|REVIEWING|COMPLETE|BLOCKED]

## Current Phase: [Phase Name]

## Completed Tasks: [ID1, ID2, ...]

## Remaining Tasks: [ID1, ID2, ...]

## Issues:
- [Issue 1]
- [Issue 2]

## Blockers:
- [Blocker description] (if STATUS: BLOCKED)

## Last Updated: [TIMESTAMP]

Quality Gates (TRAE Self-Check)

Include in implementation prompts:

SELF-CHECK before marking task complete:
- [ ] Code compiles without errors
- [ ] All tests pass
- [ ] No linting errors
- [ ] Documentation updated
- [ ] progress.md updated

Prompt Templates (Token-Efficient)

Planning

PLAN: [REQUIREMENTS]
STACK: [TECH]
OUTPUT: .trae-docs/{architecture.md, task_plan.md, progress.md}
SIGNAL: Create .trae-docs/.signal_planning_done when done

Implementation

IMPLEMENT: Tasks [IDS]
PLAN: .trae-docs/task_plan.md
ARCH: .trae-docs/architecture.md
UPDATE: .trae-docs/progress.md
SIGNAL: Create .trae-docs/.signal_batch_[N]_done when done

Review

REVIEW: All code
CHECK: .trae-docs/requirements.md
LOG: .trae-docs/review_log.md
STATUS: .trae-docs/progress.md
SIGNAL: Create .trae-docs/.signal_review_done when done

Bug Fix

FIX: [BUG_ID]
LOG: .trae-docs/review_log.md
ATTEMPTS: [N]
NEW_APPROACH: [APPROACH]
SIGNAL: Create .trae-docs/.signal_fixed_[BUG_ID] when done
OR: Create .trae-docs/.signal_blocked if still failing

Desktop Automation (Minimal)

Only needed for:

Launching TRAE
Sending initial prompt
Emergency intervention (timeout fallback)

import os
import subprocess
import pyperclip
import pyautogui

# Launch TRAE
def launch_trae(config):
    subprocess.Popen(f"{config['trae_install_path']}\\Trae CN.exe")

# Send prompt
def send_prompt(prompt_text):
    pyperclip.copy(prompt_text)
    pyautogui.hotkey('ctrl', 'v')
    pyautogui.press('enter')

# Signal file detection (0 tokens!)
def check_signal(signal_type):
    signal_path = f".trae-docs/.signal_{signal_type}"
    return os.path.exists(signal_path)

# Clean up signal after handling
def clear_signal(signal_type):
    signal_path = f".trae-docs/.signal_{signal_type}"
    if os.path.exists(signal_path):
        os.remove(signal_path)

# Main orchestration loop
def orchestrate():
    while True:
        if check_signal('planning_done'):
            progress = read_file('.trae-docs/progress.md')
            # Process and send next prompt
            clear_signal('planning_done')
            send_prompt(implementation_prompt)
            
        elif check_signal('blocked'):
            blocker = read_file('.trae-docs/progress.md')
            # Analyze and provide guidance
            clear_signal('blocked')
            send_prompt(guidance_prompt)
            
        elif check_signal('project_done'):
            # Project complete!
            break
            
        # Sleep to avoid CPU usage (no token cost)
        time.sleep(1)

Self-Update

{
  "executions": [{
    "timestamp": "ISO_DATE",
    "project": "NAME",
    "tasks": N,
    "interventions": N,
    "token_saved_estimate": N
  }]
}

Quick Reference

openclaw Action	Trigger
Check signal file	Continuous (0 tokens)
Read progress.md	Only when signal file exists
Read task_plan.md	Once per phase
Send prompt	Once per phase/batch
Intervene	Only on BLOCKED/loop

TRAE Action	Trigger
Generate code	Continuous
Create signal file	When phase done
Update progress.md	After each task
Self-check quality	After each task
Handle errors	Automatic (3 attempts)

Signal File Naming Convention

Phase	Signal File
Planning	`.signal_planning_done`
Batch N	`.signal_batch_N_done`
Review	`.signal_review_done`
Complete	`.signal_project_done`
Blocked	`.signal_blocked`

User Control Mechanism

Control Signals (User-Initiated)

User Action	Signal File	Effect
Pause	`.signal_pause`	Stop orchestration, keep TRAE running
Resume	`.signal_resume`	Continue from where paused
Stop	`.signal_stop`	Terminate project, archive progress
Skip Task	`.signal_skip_[TASK_ID]`	Skip specific task, continue next
Force Complete	`.signal_force_done`	Mark current phase as done

How to Use Control Signals

Method 1: Command Line (Windows PowerShell)

# 暂停项目
New-Item -Path ".trae-docs\.signal_pause" -ItemType file

# 恢复项目
New-Item -Path ".trae-docs\.signal_resume" -ItemType file

# 停止项目
New-Item -Path ".trae-docs\.signal_stop" -ItemType file

# 跳过任务
New-Item -Path ".trae-docs\.signal_skip_task_3" -ItemType file

# 强制完成
New-Item -Path ".trae-docs\.signal_force_done" -ItemType file

Method 2: Control Script (Recommended)

Run the control script for easy interaction:

# In project directory
python .trae/skills/trae-orchestrator/control.py

This launches an interactive menu:

TRAE Orchestrator Control Panel
================================
Current Status: RUNNING
Phase: Implementation
Progress: 5/15 tasks

[1] Pause Project
[2] Resume Project  
[3] Stop Project
[4] Skip Task
[5] Force Complete
[6] View Status
[7] Exit

Enter choice:

Method 3: Direct Python Call

from automation_helper import pause_project, resume_project, stop_project

pause_project("./my-project")   # 暂停
resume_project("./my-project")  # 恢复
stop_project("./my-project")    # 停止

Method 4: File Manager

Open project folder in file explorer
Navigate to .trae-docs/ folder
Create new text file, rename to .signal_pause (remove .txt extension)
Confirm extension change

Orchestration Loop with Control

def orchestrate(project_dir=".", handlers=None):
    while True:
        # 1. Check control signals FIRST
        if check_signal('stop', project_dir):
            archive_progress(project_dir)
            return False, "Project stopped by user"
        
        if check_signal('pause', project_dir):
            # Wait for resume signal
            while not check_signal('resume', project_dir):
                if check_signal('stop', project_dir):
                    return False, "Project stopped during pause"
                time.sleep(5)
            clear_signal('resume', project_dir)
            clear_signal('pause', project_dir)
        
        # 2. Check skip signals
        for skip_signal in get_skip_signals(project_dir):
            task_id = skip_signal.replace('skip_', '')
            mark_task_skipped(task_id, project_dir)
            clear_signal(skip_signal, project_dir)
        
        # 3. Check force complete
        if check_signal('force_done', project_dir):
            clear_signal('force_done', project_dir)
            # Move to next phase
            send_next_prompt()
        
        # 4. Normal signal processing
        signals = get_all_signals(project_dir)
        # ... rest of orchestration

Pause Behavior

When .signal_pause is detected:

┌─────────────────────────────────────────────────────────┐
│  openclaw detects .signal_pause                         │
│       ↓                                                 │
│  Stop sending new prompts                               │
│       ↓                                                 │
│  Keep TRAE running (finish current task)                │
│       ↓                                                 │
│  Wait for .signal_resume or .signal_stop                │
│       ↓                                                 │
│  Resume: Continue from last checkpoint                  │
│  Stop: Archive and terminate                            │
└─────────────────────────────────────────────────────────┘

Stop Behavior

When .signal_stop is detected:

┌─────────────────────────────────────────────────────────┐
│  openclaw detects .signal_stop                          │
│       ↓                                                 │
│  Create final progress snapshot                         │
│       ↓                                                 │
│  Archive .trae-docs/ to .trae-archive/[timestamp]/      │
│       ↓                                                 │
│  Clear all signal files                                 │
│       ↓                                                 │
│  Return control to user                                 │
└─────────────────────────────────────────────────────────┘

Status File for User Visibility

openclaw maintains .trae-docs/orchestrator_status.md:

# Orchestrator Status

## State: [RUNNING|PAUSED|STOPPED|WAITING]

## Last Action: [timestamp] - [action description]

## Next Action: [what will happen next]

## User Controls Available:
- Pause: Create .signal_pause
- Resume: Create .signal_resume (when paused)
- Stop: Create .signal_stop

## Current Progress:
- Phase: [phase name]
- Completed: N tasks
- Remaining: M tasks

Quick Commands for Users

# Check status
cat .trae-docs\orchestrator_status.md

# Or use control panel (recommended)
python .trae\skills\trae-orchestrator\control.py

# Quick commands
New-Item -Path ".trae-docs\.signal_pause" -ItemType file      # Pause
New-Item -Path ".trae-docs\.signal_resume" -ItemType file     # Resume
New-Item -Path ".trae-docs\.signal_stop" -ItemType file       # Stop
New-Item -Path ".trae-docs\.signal_skip_task_3" -ItemType file # Skip task 3
New-Item -Path ".trae-docs\.signal_force_done" -ItemType file # Force complete

Complete Workflow Example

Here's a complete example of using the automation helper:

#!/usr/bin/env python3
"""
完整示例：使用 TRAE 自动化开发一个项目
"""
from automation_helper import (
    TRAEController, 
    ProjectManager, 
    ProgressMonitor,
    quick_start,
    pause_project,
    stop_project
)

# ========== 方法 1: 一键快速启动 ==========
def method1_quick_start():
    """最简单的方式"""
    quick_start(
        project_dir='D:\\MyGame',
        requirements={
            'name': '星空篝火游戏',
            'description': '一个多人联机的3D篝火游戏',
            'features': [
                '3D星空场景',
                '多人联机',
                '聊天系统',
                '篝火效果'
            ],
            'tech_stack': 'Three.js + Node.js + Socket.io'
        },
        trae_path='E:\\software\\Trae CN\\Trae CN.exe'  # 可选，自动查找
    )

# ========== 方法 2: 分步控制 ==========
def method2_step_by_step():
    """更精细的控制"""
    
    # 1. 创建项目
    ProjectManager.create_project(
        project_dir='D:\\MyGame',
        requirements="""
# 星空篝火游戏

## 描述
创建一个多人联机的3D篝火游戏

## 功能
- 3D星空场景
- 多人联机
- 聊天系统
"""
    )
    
    # 2. 创建自定义提示
    custom_prompt = """
请开发一个星空篝火游戏。

要求：
1. 使用 Three.js 创建3D场景
2. 使用 Socket.io 实现多人联机
3. 包含星空、篝火、玩家角色
4. 实现移动、聊天、互动功能

完成后创建 .trae-docs/.signal_project_done
"""
    ProjectManager.create_prompt('D:\\MyGame', custom_prompt)
    
    # 3. 启动 TRAE
    controller = TRAEController('E:\\software\\Trae CN\\Trae CN.exe')
    controller.launch('D:\\MyGame')
    
    # 4. 发送提示
    controller.send_prompt(custom_prompt, delay=5)

# ========== 方法 3: 监控进度 ==========
def method3_monitor():
    """监控开发进度"""
    monitor = ProgressMonitor('D:\\MyGame')
    
    # 检查当前状态
    status = monitor.get_status()
    print(f"当前状态: {status}")
    
    # 等待完成（带超时）
    completed = monitor.wait_for_completion(timeout=3600)
    
    if completed:
        print("✅ 项目开发完成！")
    else:
        print("⚠️ 项目未完成或被阻塞")

# ========== 运行 ==========
if __name__ == '__main__':
    # 选择方法
    method1_quick_start()  # 最简单
    # method2_step_by_step()  # 更灵活
    # method3_monitor()  # 仅监控

File Creation Strategy

How to Teach TRAE to Create Files

Method 1: Pre-create Requirements (Recommended)

Create requirements.md BEFORE starting TRAE:

from automation_helper import ProjectManager

ProjectManager.create_project(
    project_dir='D:\\MyProject',
    requirements={
        'name': 'My App',
        'description': 'An awesome application',
        'features': ['Feature 1', 'Feature 2'],
        'tech_stack': 'React + Node.js'
    }
)

This creates:

.trae-docs/requirements.md - TRAE reads this
.trae-docs/prompt_to_trape.md - Instructions for TRAE

Then TRAE will:

Read requirements.md
Create architecture.md
Create task_plan.md
Create actual code files

Method 2: Include File List in Prompt

Create the following files:
1. src/index.js - Entry point
2. src/components/App.js - Main component
3. src/styles.css - Styles
4. package.json - Dependencies

Use this structure:

my-app/ ├── src/ │ ├── index.js │ ├── components/ │ │ └── App.js │ └── styles.css └── package.json

Method 3: Phase-Based Creation

PHASE 1 - Setup:
- Create package.json
- Create folder structure

PHASE 2 - Core:
- Create src/index.js
- Create src/app.js

PHASE 3 - UI:
- Create src/components/
- Create src/styles/

Dependencies

Required

Python 3.7+
TRAE IDE installed

Optional (for auto-send)

pip install pyautogui pyperclip

Without these, you need to manually paste the prompt into TRAE.

Troubleshooting

TRAE Not Found

from automation_helper import TRAEController

controller = TRAEController()
controller.setup('E:\\software\\Trae CN\\Trae CN.exe')  # 手动设置路径

Permission Denied

Run Python as Administrator or check TRAE path permissions.

Prompt Not Sent

Install pyautogui:

pip install pyautogui pyperclip

Or manually copy from .trae-docs/prompt_to_trape.md and paste into TRAE.

Core Principle: Event-driven orchestration. TRAE signals completion, openclaw responds. User controls via signal files. Zero polling, zero wasted tokens.

New in this version: Practical Python automation module (automation_helper.py) for one-line project launch and easy control.

TRAE指挥官

Description