Agent Workflow System

June 12, 2026 · View on GitHub

A Chinese-English bilingual AI workflow system for Claude Code and Codex CLI. Turns vague ideas into executable tasks through 7 collaborative skills + behavioral constitution + session recovery. Beginner-friendly.

中文文档 →


🚪 New Here? Start Here.

You don't need to understand how this system works. Just say what you want to do. The AI will guide you step by step — no jargon, no technical background required.


✨ Core Features

🔄 Session Recovery (v1.5.0+)

Close the window, the task survives. New window, same AI.

  • CLAUDE.md — Auto-detects unfinished tasks on session start
  • RESUME.md + context_snapshot — Checkpoint with decisions, preferences, and footgun prevention
  • Open-ended recovery — No fixed options; AI presents status, you decide direction
  • ⏱️ Time-interval awareness — Adjusts behavior based on how long since last session
  • SessionEnd Hook — Auto-saves checkpoint + commits on window close
  • SessionStart Hook — Auto git pull on session start (24h throttle)
  • Cross-platform — PowerShell + Bash

🔧 Setup (2 minutes)

Step 1: Enable hooks — Add this to your .claude/settings.local.json:

{
  "hooks": {
    "SessionStart": [{
      "matcher": "",
      "hooks": [
        {"type": "command", "command": "powershell.exe -File .claude/hooks/session-start.ps1", "timeout": 10}
      ]
    }],
    "SessionEnd": [{
      "matcher": "",
      "hooks": [
        {"type": "command", "command": "powershell.exe -File .claude/hooks/session-end.ps1", "timeout": 15}
      ]
    }]
  }
}

Hook scripts are in hooks/ (copy to .claude/hooks/). Both .ps1 and .sh versions available.

Step 2: Run health checkscripts/health-check.ps1 check for system integrity scan.

🔁 Auto-Update Pipeline

You change system → Close window → Hook auto commit + push

                                     GitHub updated

                          User opens window → Hook auto pull

🧠 7 Collaborative Skills

SkillRoleWhen
workflow-systemOrchestratorEntry routing to the right sub-skill
newbie-guideGuideVague goal → actionable task (Socratic questioning + optional background research)
project-masterPlannerTask breakdown, decision gates, acceptance criteria
debug-fixerExecutorErrors / test failures / bugs (TDD minimal fix)
learning-coachExecutorLearning programming, design, languages, etc.
drift-auditorAuditorScope creep / context drift / task fork detection
phase-closeoutCloserPhase wrap-up, freeze state, generate resume token

📜 Behavioral Constitution

15-chapter unified behavioral standard governing all skills:

  • Role isolation — Planners don't code, executors don't make architectural decisions
  • Five-level error correction — From self-check to mandatory human intervention
  • Legal awareness — Three-layer risk alerts without giving legal advice
  • Artifact handoff — 50k context → 5k artifact for lightweight inter-skill transfer
  • Session recovery — Cross-session task continuity
  • Task forking & convergence — Sub-tasks anchored to parent task context, auto-return on completion
  • Four-question self-check — Before every change: logical? conflicts? important? right timing?
  • Time-scale awareness — Tasks track created/last-active timestamps; stale decisions flagged for re-confirmation

CLAUDE.md Rules (v1.6.2 ~ v1.12.1):

  • Verify by doing — Run commands, don't guess from docs
  • Tone standards — Patient, equal, explain everything; no dismissive one-liners
  • Task closeout rule — Never close a task without user confirmation
  • Topic layer awareness — Recap what was done and where you stand in natural language
  • Four-question self-check — Pre-implementation gate: logic / conflict / importance / timing
  • Task registration rule — Every new task must be registered in task_stack before starting
  • Task fork & converge — Sub-tasks inherit parent context; auto-return on completion
  • Return-to-task format — After drift, resume with: goal / progress / next step
  • Time-scale awareness — Real-world timestamps on tasks & decisions; auto-flag stale items after idle periods

🚀 Installation

Claude Code

Option 1: Ask Claude (Recommended)

In chat:

Install agent-workflow-system for me

Option 2: Manual

git clone https://github.com/1139030773-cmd/agent-workflow-system.git
cp -r agent-workflow-system/.claude ./ && cp agent-workflow-system/CLAUDE.md . && cp agent-workflow-system/RESUME.md .

Codex

Option 1: CLI One-liner (Recommended)

codex plugin install 1139030773-cmd/agent-workflow-system

Option 2: Desktop App — Install from URL

Settings → Plugins → Menu next to search → Install from URL:

https://raw.githubusercontent.com/1139030773-cmd/agent-workflow-system/main/plugins/agent-workflow-system/.codex-plugin/plugin.json

Option 3: Marketplace (requires PR merge)

codex plugin marketplace add https://github.com/hashgraph-online/awesome-codex-plugins.git

Then type /plugins in chat and find Agent Workflow System.

💡 Marketplace install depends on PR #202. Use Option 1 or 2 in the meantime.

Starter Prompts

After installation, try one of these:

ScenarioSay
🆕 Vague idea"Start the newbie guide, help me turn my idea into tasks"
📋 Manage a project"Enter project master mode, help me manage this project"
🐛 Something's broken"Help me fix this bug"
📖 Learn something"I want to learn Python, make me a study plan"
🔍 Feeling off-track"Check if my progress has drifted"
📦 Wrap up"Help me do a phase closeout and organize what we did"

💡 No commands to memorize. Speak naturally. The system auto-detects your language and intent.


📖 Version History

VersionDateHighlights
1.12.12026-06-12🛡️ Adversarial verification — independent agent checks for "should've changed but didn't" before release
1.12.02026-06-12🔍 Research gate — optional background research between newbie-guide and project-master
1.11.12026-06-12health-check.ps1 encoding fix — 3 bugs, 100% health
1.11.02026-06-12PROJECT.md — one-screen overview, auto-updated per phase
1.10.32026-06-11Removed write-level gate — classification interference
1.10.22026-06-11Full distribution — scripts + hooks included in plugin package
1.10.02026-06-11One-click project archive — INDEX.md + NL recall
1.9.22026-06-11Output self-check Q4 + phase-transition auto audit
1.9.12026-06-11Project dashboard — health_score + active_tasks in RESUME.md
1.9.02026-06-11Beginner mode — newbie entry + jargon filter + beginner channel
1.8.22026-06-11Output self-check + write-level gate — prevent dev content leaking into workflow
1.8.12026-06-11System health check + auto-cleanup — integrity scan on every session start
1.8.02026-06-11Multi-window session management — .resume/ per-window files prevent conflicts
1.7.72026-06-10Time-scale awareness — task timestamps + decision expiry + idle detection
1.7.62026-06-10English adaptation — bilingual README + language auto-detect + starter prompts
1.7.52026-06-09Skill exit protocol + Return-to-task format rule
1.7.42026-06-09Mandatory format for drift → task return
1.7.32026-06-09Sub-task anchoring: parent context = practice material
1.7.22026-06-09Task registration rule + background recap on return
1.7.12026-06-07Session closeout reminder + Version consistency validator
1.7.02026-06-07CLAUDE.md split (dev rules / user rules)
1.6.32026-06-07Task closeout rule
1.6.22026-06-05Verify by doing + Tone standards
1.6.12026-06-05Time-interval awareness + 7-step change protocol + Auto-update pipeline
1.6.02026-06-05context_snapshot + Single-thread principle
1.5.02026-06-05Session recovery mechanism, cross-platform hooks

See CHANGELOG.md for full details.


📄 License

MIT