Agent Workflow System：为 Claude Code 和 Codex 打造的双语 AI 协作工作流系统

Core Overview

Agent Workflow System is an open-source bilingual (Chinese/English) AI agent workflow system developed by 1139030773-cmd and released on GitHub in June 2026. It aims to convert vague ideas into executable tasks via three core mechanisms:

1. 7 specialized collaboration roles forming a PDCA loop
2. 14-chapter behavior规范章程 (behavior norms)
3. Session recovery for cross-session continuity It supports both Claude Code and Codex CLI platforms, serving as an efficient AI collaboration partner for developers.

Source: GitHub Repository

With the popularity of AI programming assistants like Claude Code and Codex CLI, developers face key challenges:

• Converting vague ideas into AI-understandable explicit tasks
• Maintaining context coherence in multi-round dialogues
• Preventing task drift or scope creep
• Losing context when closing conversation windows Agent Workflow System is designed to address these issues by introducing structured collaboration mechanisms.

7 Collaboration Roles

The system decomposes AI functions into 7 roles forming a closed loop:

1. workflow-system: Coordinator (route user intent to appropriate skills)
2. newbie-guide: Newbie mentor (Socratic questioning to clarify vague goals)
3. project-master: Project planner (task breakdown, checkpoint setting)
4. debug-fixer: Debugger (TDD-based bug fixing)
5. learning-coach: Learning guide (structured learning paths)
6. drift-auditor: Drift monitor (alert on scope creep/context drift)
7. phase-closeout: Phase finisher (summary, state freezing, recovery token generation)

14-Chapter Behavior Norms

Key norms include:

• Role isolation (planners don’t code, executors don’t make architecture decisions)
• 5-level error correction mechanism
• Legal risk预警 (3 layers)
• Artifact handover optimization (50k context →5k lightweight artifacts)
• Task fork/convergence (subtasks inherit parent context)
• 4-question self-check before changes (logic, conflict, importance, timing)

Traditional AI assistants lose context when windows close. This system solves it via:

1. CLAUDE.md auto-detection: Check for unfinished tasks at session start
2. RESUME.md + context_snapshot: Save decision records, user preferences, and notes as checkpoints
3. Time interval awareness: Adjust recovery strategy based on time since last session (e.g., detailed回顾 for >24h gaps)
4. SessionEnd hook: Auto-save checkpoints and commit changes when window closes
5. SessionStart hook: Auto git pull (24h throttle) for code sync
6. Cross-platform support: Works with PowerShell and Bash

Claude Code Installation

• Natural language: Say "Install agent-workflow-system for me" to Claude
• Manual: Clone the repository and copy config files

Codex CLI Installation

• Command line: codex plugin install 1139030773-cmd/agent-workflow-system
• Desktop app: Install via plugin market

The system supports intuitive natural language启动 without complex commands:

Scenario	Launch Command
Vague idea to tasks	"Start the newbie guide, help me turn my idea into tasks"
Project management	"Enter project master mode, help me manage this project"
Bug fixing	"Help me fix this bug"
Learning new skills	"I want to learn Python, make me a study plan"
Check task drift	"Check if my progress has drifted"
Phase summary	"Help me do a phase closeout and organize what we did"

It auto-detects Chinese/English input and adapts accordingly.

Version Updates

• v1.5.0: Introduced session recovery mechanism
• v1.7.5: Added skill exit protocol and task return format规范

Community & License

• License: MIT (open source)
• Status: Submitted to Codex community plugin market (waiting PR merge)
• Active update rhythm addressing real user pain points

Agent Workflow System represents a new AI collaboration paradigm: treating AI as a professional partner with role division, consistent norms, and memory. It ensures:

• Professionalism via role separation
• Consistency via behavior norms
• Continuity via session recovery

For developers using Claude Code or Codex CLI, this system is worth trying—it may transform how you collaborate with AI, turning vague ideas into executable code.