A complete development toolchain for AI startup teams: 6 Development Workflow Skills + Codex Autopilot Multi-Project Engine + OpenClaw / Hermes Intelligent Orchestration Layer.

Requirement → Doc Writing → Doc Review → Development ←→ Testing → Code Review → Release

Core Principles: Startup-friendly (MoSCoW MVP) · AI-native (token cost controls) · SOLID-driven · Doc-closed-loop

# Gemini
ln -sf /path/to/AIWorkFlowSkill/development ~/.gemini/skills/development

# Codex — reference in AGENTS.md
# Claude — add to Skills directory

  Trigger          Detection         Decision          Execution
┌──────────┐    ┌─────────────┐   ┌─────────────┐   ┌─────────────┐
│ launchd  │───→│codex-status │──→│ watchdog.sh │──→│ tmux-send.sh│
│  (10s)   │    │    .sh      │   │ (~1700 LOC) │   │ (3-layer)   │
└──────────┘    │ JSON output │   │ State machine│   └─────────────┘
│  cron    │───→│working/idle/│   │ Exp backoff  │──→│ task-queue  │
│ (10min)  │    │perm/shell   │   │ Lock/compact │   │    .sh      │
└──────────┘    └─────────────┘   └─────────────┘   └─────────────┘

User (Telegram) → "Fix white-screen bug"
  → Claude (OpenClaw) writes to task-queue
  → Watchdog detects Codex idle → dispatches task
  → Codex fixes → commit → triggers Claude code review
  → Review clean → Discord notification "✅ Bug fixed"

Task Queue
├─ type: frontend/ui/h5  → 🎨 Gemini tmux window (design-optimized)
├─ type: bugfix/feature   → 🔧 Codex tmux window (code-optimized)
├─ Codex limit exhausted  → 🤖 Claude AgentTeam fallback
└─ Gemini unavailable     → 🔧 Codex fallback (graceful degradation)

Frontend tasks do not go through ACP for now. Direct tmux persistent sessions are more stable in current production runs.

Configuration:

# config.yaml
gemini:
  default_window: "gemini-h5"    # Default Gemini tmux window
  project_windows:               # Per-project mapping
    youxin: "gemini-youxin"

Usage:

# Frontend tasks route to Gemini automatically
task-queue.sh add myproject "Build login page" normal --type frontend

# Backend tasks still go to Codex
task-queue.sh add myproject "Fix auth API" high --type bugfix

Frontend tasks include Anti-AI-Slop prompt injection: layout checks, design system consistency, interaction state coverage (loading/empty/error/success).

Trigger points:

• on_commit_evaluate: evaluate test/coverage right after commit detection
• on_review_clean: run coverage-gap analysis and enqueue test tasks after review is clean
• nightly: scheduled coverage evaluation window

Flow:

commit
  → watchdog (detect new commit)
  → test-agent evaluate (run tests + collect coverage)
  → pass: continue workflow, then on review clean run coverage-gap enqueue
  → fail: auto-parse test logs
          → extract failed test files + error summary
          → task-queue add "fix tests" bugfix tasks (high priority)

Coverage ratchet policy:

• Weekly ratchet: +1%
• Cap: 90%

Auto-enqueue on failure (introduced in 386a682):

• Parse $HOME/.autopilot/logs/test-agent-run-*.log and package run logs
• Extract failed test file and key error line
• Enqueue bugfix task via task-queue.sh add <project> ... high --type bugfix
• 1-hour cooldown dedupe per failed target to avoid retry loops

Config example:

test_agent:
  enabled: true
  trigger:
    on_commit_evaluate: true
    on_review_clean: true
    nightly: "02:30"
  queue:
    max_tasks_per_round: 3
  coverage:
    changed_files_min: 80
    ratchet_weekly: 1
    ratchet_cap: 90

Codex idle
├─ Queue has tasks? → consume queue (bypass cooldown)
│   ├─ type=frontend? → route to Gemini window
│   └─ type=other?    → route to Codex window
├─ Review has issues? → nudge #N/5 (5-attempt cap, backoff)
├─ Compact just finished? → resume with context snapshot
├─ PRD has issues? → nudge fix
├─ Nothing pending → 💤 stay quiet (zero token waste)
└─ Dirty tree? → prompt commit

For unattended Codex tmux sessions, the default production setup is full permissions:

echo "alias codex='command codex --yolo'" >> ~/.zshrc
source ~/.zshrc

If you do not want a shell alias, invoke command codex --yolo explicitly inside the tmux window that Autopilot manages.

Use the operational helpers below to close the loop around task dispatch:

# Wait until a dispatched Codex session is idle again
./scripts/watch-codex.sh autopilot:ProjectA 60

# Check whether upstream PRs received new reviews/comments
./scripts/pr-monitor.sh

# Snapshot local workspace files and memory to a GitHub backup repo
./scripts/gitclaw-backup.sh

• watch-codex.sh is the lightweight tmux-side health check used after dispatch when you need an exact "idle again" signal.
• pr-monitor.sh detects fresh review activity and merge/close events so OpenClaw or Hermes agents can enqueue follow-up or CI-fix tasks.
• docs/HEARTBEAT.md is included as an example production heartbeat checklist for recurring PR follow-up, growth checks, and maintenance loops.

# 1. Configure projects
cat > watchdog-projects.conf << EOF
ProjectA:/path/to/project-a:Default nudge message
EOF

# 2. Configure Telegram (config.yaml)
# 3. Recommended: make Codex default to full permissions
echo "alias codex='command codex --yolo'" >> ~/.zshrc
source ~/.zshrc

# 4. Create tmux session + start Codex
tmux new-session -s autopilot -n ProjectA
codex

# 5. Start watchdog
nohup bash scripts/watchdog.sh &

# 6. Submit tasks
bash scripts/task-queue.sh add myproject "Fix bug" high

AIWorkFlowSkill/
├── README.md
├── CONVENTIONS.md
├── CONTRIBUTING.md
├── LICENSE
├── requirement-discovery/    # Skill: Requirement Research
├── doc-writing/              # Skill: Doc Writing
├── doc-review/               # Skill: Doc Review
├── development/              # Skill: Development
├── testing/                  # Skill: Test Design
├── code-review/              # Skill: Code Review
├── scripts/                  # Autopilot Engine
│   ├── watchdog.sh
│   ├── codex-status.sh
│   ├── tmux-send.sh
│   ├── monitor-all.sh
│   ├── task-queue.sh
│   ├── watch-codex.sh
│   ├── pr-monitor.sh
│   ├── gitclaw-backup.sh
│   └── ...
├── docs/
│   └── HEARTBEAT.md
├── watchdog-projects.conf
├── config.yaml
└── prd-items.yaml

MIT

Built on OpenClaw, Hermes-style orchestration patterns, and Codex CLI.