turix-mac（turix Mac 搜索）

TuriX-Mac Skill

This skill allows Clawdbot to control the macOS desktop visually using the TuriX Computer Use Agent.

When to Use

• When asked to perform actions on the Mac desktop (e.g., "Open Spotify and play my liked songs").
• When navigating applications that lack command-line interfaces.
• For multi-step visual workflows (e.g., "Find the latest invoice in my email and upload it to the company portal").
• When you need the agent to plan, reason, and execute complex tasks autonomously.

Key Features

🤖 Multi-Model Architecture

TuriX uses a sophisticated multi-model system:

• Brain: Understands the task and generates step-by-step plans
• Actor: Executes precise UI actions based on visual understanding
• Planner: Coordinates high-level task decomposition (when use_plan: true)
• Memory: Maintains context across task steps

📋 Skills System

Skills are markdown playbooks that guide the agent for specific domains:

• github-web-actions: GitHub navigation, repo search, starring
• browser-tasks: General web browser operations
• Custom skills can be added to the skills/ directory

🔄 Resume Capability

The agent can resume interrupted tasks by setting a stable agent_id.

Running TuriX

Basic Task

skills/local/turix-mac/scripts/run_turix.sh "Open Chrome and go to github.com"

Resume Interrupted Task

skills/local/turix-mac/scripts/run_turix.sh --resume my-task-001

✅ Note: run_turix.sh updates examples/config.json for you (task, resume, use_plan, use_skills). If you want to keep a hand-edited config, skip passing a task and edit examples/config.json directly.

Tips for Effective Tasks

✅ Good Examples:

• "Open Safari, go to google.com, search for 'TuriX AI', and click the first result"
• "Open System Settings, click on Dark Mode, then return to System Settings"
• "Open Finder, navigate to Documents, and create a new folder named 'Project X'"

❌ Avoid:

• Vague instructions: "Help me" or "Fix this"
• Impossible actions: "Delete all files"
• Tasks requiring system-level permissions without warning

💡 Best Practices:

1. Be specific about the target application
2. Break complex tasks into clear steps, but do not mention the precise coordinates on the screen.

Hotkeys

• Force Stop: Cmd+Shift+2 - Immediately stops the agent

Monitoring & Logs

Logs are saved to .turix_tmp/logging.log in the project directory. Check this for:

• Step-by-step execution details
• LLM interactions and reasoning
• Errors and recovery attempts

Important Notes

How TuriX Runs

• TuriX can be started via clawdbot exec with pty:true mode
• The first launch takes 2-5 minutes to load all AI models (Brain, Actor, Planner, Memory)
• Background output is buffered - you won't see live progress until task completes or stops

Before Running

Always set PATH first:

export PATH="/usr/sbin:$PATH"
cd your_dir/TuriX-CUA
/opt/anaconda3/envs/turix_env/bin/python examples/main.py

Why? The screencapture tool is located at /usr/sbin/screencapture, which is not in the default PATH.

Checking if TuriX is Running

# Check process
ps aux | grep "python.*main" | grep -v grep

# Should show something like:
# user  57425  0.0  2.4 412396704 600496 s143  Ss+  5:56PM   0:04.76 /opt/anaconda3/envs/turix_env/bin/python examples/main.py

Note: The .turix_tmp directory may not be created until TuriX starts executing steps.

Troubleshooting

Common Issues

Debug Mode

Logs include DEBUG level by default. Check:

tail -f your_dir/TuriX-CUA/.turix_tmp/logging.log

Architecture

User Request
     ↓
[Clawdbot] → [TuriX Skill] → [run_turix.sh] → [TuriX Agent]
                                              ↓
                    ┌─────────────────────────┼─────────────────────────┐
                    ↓                         ↓                         ↓
               [Planner]                 [Brain]                  [Memory]
                    ↓                         ↓                         ↓
                                         [Actor] ───→ [Controller] ───→ [macOS UI]

Skill System Details

Skills are markdown files with YAML frontmatter in the skills/ directory:

---
name: skill-name
description: When to use this skill
---
# Skill Instructions
High-level workflow like: Open Safari,then go to Google.

The Planner selects relevant skills based on name/description; the Brain uses full content for step guidance.

Advanced Options

TuriX Skills System

TuriX supports Skills: markdown playbooks that help the agent behave more reliably in specific domains.

1. Built-in Skills

2. Create a Custom Skill

Create a .md file in the TuriX project's skills/ directory:

---
name: my-custom-skill
description: When performing X specific task
---
# Custom Skill

## Guidelines
- Step 1: Do this first
- Step 2: Then do that
- Step 3: Verify the result

Field definitions:

• name: Skill identifier (used by the Planner to select)
• description: When to use this skill (Planner matches on this)
• The body below: Full execution guide (used by the Brain)

3. Enable Skills

In examples/config.json:

{
  "agent": {
    "use_plan": true,
    "use_skills": true,
    "skills_dir": "skills",
    "skills_max_chars": 4000
  }
}

4. Run a Task with Skills

skills/local/turix-mac/scripts/run_turix.sh "Search for turix-cua on GitHub and star it"

The agent will automatically:

1. Planner reads the skill name and description
2. Selects relevant skills
3. Brain uses the full skill content to guide execution

5. Chinese Text Support

Background: Passing Chinese text through shell interpolation can mangle UTF-8, and interpolating untrusted text into a heredoc is unsafe.

Solution: The run_turix.sh script uses Python to handle UTF-8 correctly and reads task text from environment variables:

import json

# Read with UTF-8
with open(config_path, 'r', encoding='utf-8') as f:
    data = json.load(f)

# Write without escaping non-ASCII text
with open(config_path, 'w', encoding='utf-8') as f:
    json.dump(data, f, indent=2, ensure_ascii=False)

Key points:

1. Always use encoding='utf-8' when reading/writing files
2. Use ensure_ascii=False to preserve non-ASCII text
3. Pass task content via environment variables or stdin, and use a single-quoted heredoc to avoid shell interpolation

6. Document Creation Best Practices

Challenges:

• Asking TuriX to collect news, then create and send a document directly
• TuriX is a GUI agent, so it can be slow and less deterministic. Prefer using TuriX only for tasks Clawdbot cannot do or where TuriX is faster.

Recommended approach: create the document yourself and let TuriX only send it

1. Create the Word document with python-docx
2. Let TuriX only send the file

from docx import Document
doc = Document()
doc.add_heading('Title')
doc.save('/path/to/file.docx')

Suggested workflow:

1. Use web_fetch to gather information
2. Use Python to create the Word document
3. Use TuriX to send the file. Specify the file path and say to send the file, not just the file name.
4. If you really need TuriX to manually create a Word document and type in collected information, put the content in turix skills (for large amounts) or in the task name (for small amounts).

7. Example: Add a New Skill

Create skills/browser-tasks.md:

---
name: browser-tasks
description: When performing tasks in a web browser (search, navigate, fill forms).
---
# Browser Tasks

## Navigation
- Use the address bar or search box to navigate
- Open new tabs for each distinct task
- Wait for page to fully load before proceeding

## Forms
- Click on input fields to focus
- Type content clearly
- Look for submit/button to complete actions

## Safety
- Confirm before submitting forms
- Do not download files without user permission

8. Skill Development Tips

1. Be precise in the description - helps the Planner select correctly
2. Make steps clear - the Brain needs explicit guidance
3. Include safety checks - confirmations for important actions
4. Keep it concise - recommended under 4000 characters

Monitoring and Debugging Guide

1. Run a Task

# Run in background (recommended)
cd your_dir/clawd/skills/local/turix-mac/scripts
./run_turix.sh "Your task description" --background

# Or use timeout to set a max runtime
./run_turix.sh "Task" &

2. Monitor Progress

Method 1: Session logs

# List running sessions
clawdbot sessions_list

# View history
clawdbot sessions_history <session_key>

Method 2: TuriX logs

# Tail logs in real time
tail -f your_dir/TuriX-CUA/.turix_tmp/logging.log

# Or inspect completed step files
ls -lt your_dir/TuriX-CUA/examples/.turix_tmp/brain_llm_interactions.log_brain_*.txt

Method 3: Check processes

ps aux | grep "python.*main.py" | grep -v grep

Method 4: Check generated files

# List files created by the agent
ls -la your_dir/TuriX-CUA/examples/.turix_tmp/*.txt

3. Log File Reference

Key log markers:

• 📍 Step N - New step started
• ✅ Eval: Success/Failed - Current step evaluation
• 🎯 Goal to achieve this step - Current goal
• 🛠️ Action - Executed action
• ✅ Task completed successfully - Task completed

4. Common Monitoring Issues

5. Force Stop

Hotkey: Cmd+Shift+2 - stop the agent immediately

Command:

pkill -f "python examples/main.py"

6. View Results

After completion, the agent will:

1. Create interaction logs in .turix_tmp/
2. Create record files (if record_info is used)
3. Keep screenshots in memory for subsequent steps

Example: view a summary file

cat your_dir/TuriX-CUA/examples/.turix_tmp/latest_ai_news_summary_jan2026.txt

7. Debugging Tips

1. Inspect Brain reasoning: check brain_llm_interactions.log_brain_*.txt for analysis and next_goal
2. Inspect Actor actions: check actor_llm_interactions.log_actor_*.txt for actions
3. Check screenshots: TuriX captures a screenshot each step (kept in memory)
4. Read record files: the agent uses record_info to save key info to .txt files

8. Example Monitoring Flow

# 1. Run a task
./run_turix.sh "Search AI news and summarize" &

# 2. Wait a few seconds and check the process
sleep 10 && ps aux | grep main.py

# 3. Check if logs are being created
ls -la your_dir/TuriX-CUA/examples/.turix_tmp/

# 4. Tail progress in real time
tail -f your_dir/TuriX-CUA/.turix_tmp/logging.log

# 5. Check current step count
ls your_dir/TuriX-CUA/examples/.turix_tmp/brain_llm_interactions.log_brain_*.txt | wc -l