cyanluna-git/cyanluna.skills

cyanluna.skills

AI-powered kanban pipeline for Claude Code and Codex — seven autonomous agents, one board.

---

Click to watch demo (25s)

---

Quick Start

1. Clone and install skills

git clone https://github.com/cyanluna/cyanluna.skills.git
REPO="$PWD/cyanluna.skills"

# Claude install
mkdir -p ~/.claude/skills
cp -R "$REPO"/kanban         ~/.claude/skills/
cp -R "$REPO"/kanban-run     ~/.claude/skills/
cp -R "$REPO"/kanban-refine  ~/.claude/skills/
cp -R "$REPO"/kanban-init    ~/.claude/skills/
cp -R "$REPO"/kanban-explore ~/.claude/skills/
cp -R "$REPO"/kanban-board   ~/.claude/kanban-board

# Codex install (recommended: symlink to shared source)
mkdir -p ~/.codex/skills
for s in kanban kanban-run kanban-refine kanban-init kanban-explore; do
  ln -sfn "$REPO/$s" "$HOME/.codex/skills/$s"
done
ln -sfn "$HOME/.claude/kanban-board" "$HOME/.codex/kanban-board"

2. Set up environment

cd ~/.codex/kanban-board 2>/dev/null || cd ~/.claude/kanban-board
cp .env.example .env   # fill in DATABASE_URL and optionally Cloudflare R2 vars
pnpm install

Create a free Neon database at neon.tech and paste the connection string as DATABASE_URL. For image attachments, set up a Cloudflare R2 bucket and fill in the CLOUDFLARE_R2_* variables. Image uploads are disabled if these are omitted.

3. Initialize a project (inside any project directory)

/kanban-init

This creates .claude/kanban.json, .codex/kanban.json, and a kanban-board/start.sh launcher. Project data is stored in Neon under a project column — no local DB files needed.

4. Start the board and add tasks

./kanban-board/start.sh        # opens http://localhost:5173

/kanban add Implement user authentication
/kanban-run 1                  # runs the full AI pipeline

---

The Pipeline

Every task flows through a 7-column board. AI agents handle each stage automatically.

Req → Plan → Review Plan → Impl → Review Impl → Test → Done

Model routing is provider-aware via kanban/models.json. Under Codex, kanban-run and kanban-batch-run intentionally resolve to the higher-capability Codex route for the pipeline agents.

Pipeline Levels

Not every task needs the full pipeline. Set the level at creation time:

---

The AI Team

Each agent has a fixed nickname used as a signature in every field and log entry. The task card becomes a complete work record — you can always see who wrote what and when.

Signature rule — every agent prepends a header to its output:

> **Planner** `opus` · 2026-02-24T10:00:00Z

Scoring Rubrics

Review agents use structured scoring (1–5 per dimension) instead of plain approve/reject.

Critic scores plans on 3 dimensions:

Average >= 4.0 → approved. Any score = 1 or average < 3.0 → changes requested.

Inspector scores implementations on 7 dimensions:

Completion = 1, Security = 1, or Type Safety = 1 → hard reject.

Done-When Verification Chain

The done_when field connects agents into a verification loop:

1. Planner writes a done_when checklist with verifiable completion criteria
2. Critic reviews done_when quality — low score triggers /kanban-refine recommendation
3. Builder must verify every done_when item before finishing
4. Inspector checks that all done_when criteria are actually met

---

Web Board Features

Three views accessible from the top tab bar:

Board

• 7-column kanban with real-time task counts
• Drag-and-drop between columns (enforces valid status transitions)
• Card detail modal with lifecycle progress bar, editable requirements, level selector

List

• Inline editing of status, level, and priority without opening the modal

Chronicle (연대기)

• Timeline view of every lifecycle event across all tasks, grouped by ISO week
• 6 event types: Created · Started · Plan ready · Reviewed · Tested · Completed
• Agent Activity toggle — expands agent_log entries into the timeline
• Click any event row to open the full task detail modal

Common

• Search by title, description, tags, or #ID
• Sort by creation date, completion date, or default rank (persisted in localStorage)
• Hide old Done toggle (3d+ threshold, persisted in localStorage)
• Multi-project support — all projects on one board, or filter by project (persisted in localStorage)
• Copy card reference — click to copy #ID Title to clipboard
• Notes with markdown support
• Image attachments with drag-and-drop upload (stored in Cloudflare R2)
• Markdown rendering in plan, implementation notes, and reviews
• Mermaid diagrams rendered inline
• Agent log viewer — full chronological history of all agents per task
• 10s auto-refresh (pauses when modal is open or dragging)
• Dark theme by default

---

Commands Reference

---

Architecture

~/.claude/ and ~/.codex/
├── skills/
│   ├── kanban/              # CRUD & board (SKILL.md + shared context + schema + templates)
│   │   ├── SKILL.md
│   │   ├── shared.md        # Shared context (pipeline, API endpoints, error handling)
│   │   ├── schema.md        # PostgreSQL schema & JSON field formats
│   │   └── templates/       # Agent prompt templates
│   │       ├── plan-agent.md
│   │       ├── review-agent.md
│   │       ├── worker-agent.md
│   │       ├── tdd-tester.md
│   │       ├── code-review-agent.md
│   │       └── test-runner.md
│   ├── kanban-run/          # Pipeline orchestration
│   │   └── SKILL.md
│   ├── kanban-refine/       # Requirements refinement interview
│   │   └── SKILL.md
│   ├── kanban-explore/      # Codebase exploration & task seeding
│   │   └── SKILL.md
│   └── kanban-init/         # Project registration skill
│       └── SKILL.md
└── kanban-board/            # Central web board (Vite + TypeScript → Neon PostgreSQL)
    ├── plugins/kanban-api.ts
    ├── .env.example         # Template — copy to .env and fill in credentials
    └── .env                 # DATABASE_URL + CLOUDFLARE_R2_* (gitignored)

Neon PostgreSQL              # Centralized DB — all projects, all PCs
└── tasks table              # `project` column isolates per-project data

/
├── .claude/kanban.json      # Project config {"project": "my-project"}
├── .codex/kanban.json       # Same project config for Codex
└── kanban-board/start.sh    # Launcher: ~/.codex/kanban-board 우선, 없으면 ~/.claude/kanban-board

All task data lives in Neon — accessible from any machine without file sync.

---

Cross-PC Sync

Task data is stored in Neon PostgreSQL — sync across PCs is built-in. Any machine with the DATABASE_URL and the kanban-board running sees the same data instantly.

No OneDrive, no symlinks, no WAL conflicts.

---

Other Skills

This repo also includes utility skills:

Install: cp -R ~/.claude/skills/ or ~/.codex/skills/ (or symlink from repo)

---

License

MIT