One Karo (manager) coordinating 7 Ashigaru (workers) + 1 Gunshi (strategist) — real session, no mock data.

Quick Start

Requirements: tmux, bash 4+, at least one of: Claude Code / Codex / Copilot / Kimi

git clone https://github.com/yohey-w/multi-agent-shogun
cd multi-agent-shogun
bash first_setup.sh                        # one-time setup: config, dependencies, MCP
source ~/.bashrc                           # reload PATH
claude --dangerously-skip-permissions      # first run only: OAuth + accept Bypass Permissions → /exit
bash shutsujin_departure.sh                # launch all agents

For full install steps (incl. Windows) and the first-30-minutes walkthrough, see 🚀 Quick Start and the basic usage section below.

Type a command in the Shogun pane:

"Build a REST API for user authentication"

Shogun delegates → Karo breaks it down → 7 Ashigaru execute in parallel. You watch the dashboard. That's it.

Want to go deeper? The rest of this README covers architecture, configuration, memory design, and multi-CLI setup.

What is this?

multi-agent-shogun is a system that runs multiple AI coding CLI instances simultaneously, orchestrating them like a feudal Japanese army. Supports Claude Code, OpenAI Codex, GitHub Copilot, and Kimi Code.

Why use it?

• One command spawns 7 AI workers + 1 strategist executing in parallel
• Zero wait time — give your next order while tasks run in the background
• AI remembers your preferences across sessions (Memory MCP)
• Real-time progress on a dashboard

        You (上様 / The Lord)
             │
             ▼  Give orders
      ┌─────────────┐
      │   SHOGUN    │  ← Receives your command, delegates instantly
      └──────┬──────┘
             │  YAML + tmux
      ┌──────▼──────┐
      │    KARO     │  ← Distributes tasks to workers
      └──────┬──────┘
             │
    ┌─┬─┬─┬─┴─┬─┬─┬─┬────────┐
    │1│2│3│4│5│6│7│ GUNSHI │  ← 7 workers + 1 strategist
    └─┴─┴─┴─┴─┴─┴─┴────────┘
       ASHIGARU      軍師

Why Shogun?

Most multi-agent frameworks burn API tokens on coordination. Shogun doesn't.

What makes this different

Zero coordination overhead — Agents talk through YAML files on disk. The only API calls are for actual work, not orchestration. Run 8 agents and pay only for 8 agents' work.

Full transparency — Every agent runs in a visible tmux pane. Every instruction, report, and decision is a plain YAML file you can read, diff, and version-control. No black boxes.

Battle-tested hierarchy — The Shogun → Karo → Ashigaru chain of command prevents conflicts by design: clear ownership, dedicated files per agent, event-driven communication, no polling.

Why CLI (Not API)?

Most AI coding tools charge per token. Running 8 Opus-grade agents through the API costs $100+/hour. CLI subscriptions flip this:

"Use AI recklessly" — With flat-rate CLI subscriptions, deploy 8 agents without hesitation. The cost is the same whether they work 1 hour or 24 hours. No more choosing between "good enough" and "thorough" — just run more agents.

Multi-CLI Support

Shogun isn't locked to one vendor. The system supports 4 CLI tools, each with unique strengths:

A unified instruction build system generates CLI-specific instruction files from shared templates:

instructions/
├── common/              # Shared rules (all CLIs)
├── cli_specific/        # CLI-specific tool descriptions
│   ├── claude_tools.md  # Claude Code tools & features
│   └── copilot_tools.md # GitHub Copilot CLI tools & features
└── roles/               # Role definitions (shogun, karo, ashigaru)
    ↓ build
CLAUDE.md / AGENTS.md / copilot-instructions.md  ← Generated per CLI

One source of truth, zero sync drift. Change a rule once, all CLIs get it.

Bottom-Up Skill Discovery

This is the feature no other framework has.

As Ashigaru execute tasks, they automatically identify reusable patterns and propose them as skill candidates. The Karo aggregates these proposals in dashboard.md, and you — the Lord — decide what gets promoted to a permanent skill.

Ashigaru finishes a task
    ↓
Notices: "I've done this pattern 3 times across different projects"
    ↓
Reports in YAML:  skill_candidate:
                     found: true
                     name: "api-endpoint-scaffold"
                     reason: "Same REST scaffold pattern used in 3 projects"
    ↓
Appears in dashboard.md → You approve → Skill created in .claude/commands/
    ↓
Any agent can now invoke /api-endpoint-scaffold

Skills grow organically from real work — not from a predefined template library. Your skill set becomes a reflection of your workflow.

Quick Start

Windows (WSL2)

cd /mnt/c/tools/multi-agent-shogun
./first_setup.sh

./shutsujin_departure.sh

First-time only: Authentication

After first_setup.sh, run these commands once to authenticate:

# 1. Apply PATH changes
source ~/.bashrc

# 2. OAuth login + Bypass Permissions approval (one command)
claude --dangerously-skip-permissions
#    → Browser opens → Log in with Anthropic account → Return to CLI
#    → "Bypass Permissions" prompt appears → Select "Yes, I accept" (↓ to option 2, Enter)
#    → Type /exit to quit

This saves credentials to ~/.claude/ — you won't need to do it again.

Daily startup

Open an Ubuntu terminal (WSL) and run:

cd /mnt/c/tools/multi-agent-shogun
./shutsujin_departure.sh

📱 Mobile Access — Dedicated Android App (Recommended)

Monitor and command 10 AI agents from your phone with the dedicated Android companion app.

Note: Android only for now. No iOS version — the developer doesn't own an iPhone. If there's demand, please open an Issue. PRs welcome!

Setup

Prerequisites:

• Shogun system running on WSL2 (or Linux server)
• SSH server started (sudo service ssh start)
• Phone and server on same network (LAN or Tailscale)

Steps:

1. Install APK

   1. Download android/release/multi-agent-shogun.apk on your phone (open the file on GitHub → "Download raw file")
   2. Tap the download notification → "Install"
   3. If "Unknown sources" warning appears → "Settings" → enable "Allow from this source" for your browser → go back → "Install"
   4. Done → "Open"

2. Configure SSH (Settings tab)

   Field	Example	Description
   SSH Host	100.xxx.xxx.xxx	Server IP (e.g., Tailscale IP)
   SSH Port	22	Usually 22
   SSH User	your_username	SSH login username
   SSH Key Path	/data/data/.../id_ed25519	Private key path on phone (*1)
   SSH Password	****	Use if no key available
   Project Path	/mnt/c/tools/multi-agent-shogun	Server-side project directory
   Shogun Session	shogun	tmux session name for Shogun
   Agent Session	multiagent	tmux session name for agents

   *1 Transfer your private key to the phone, or use password authentication

3. Save → Switch to Shogun tab → auto-connects

Using Tailscale (connect from anywhere):

# Server-side (WSL2)
curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscaled &
sudo tailscale up --authkey tskey-auth-XXXXXXXXXXXX
sudo service ssh start

Install the Tailscale app on your phone, log in with the same account, and use the displayed Tailscale IP as the SSH Host in the app.

With ntfy notifications:

See ntfy setup section for push notifications from Karo on task completion.

# 1. Clone
git clone https://github.com/yohey-w/multi-agent-shogun.git ~/multi-agent-shogun
cd ~/multi-agent-shogun

# 2. Make scripts executable
chmod +x *.sh

# 3. Run first-time setup
./first_setup.sh

cd ~/multi-agent-shogun
./shutsujin_departure.sh

wsl --install

After Setup

Whichever option you chose, 10 AI agents are automatically launched:

Two tmux sessions are created:

• shogun — connect here to give commands
• multiagent — Karo, Ashigaru, and Gunshi running in the background

How It Works

Step 1: Connect to the Shogun

After running shutsujin_departure.sh, all agents automatically load their instructions and are ready.

Open a new terminal and connect:

tmux attach-session -t shogun

Step 2: Give your first order

The Shogun is already initialized — just give a command:

Research the top 5 JavaScript frameworks and create a comparison table

The Shogun will:

1. Write the task to a YAML file
2. Notify the Karo (manager)
3. Return control to you immediately — no waiting!

Meanwhile, the Karo distributes tasks to Ashigaru workers for parallel execution.

Step 3: Check progress

Open dashboard.md in your editor for a real-time status view:

## In Progress
| Worker | Task | Status |
|--------|------|--------|
| Ashigaru 1 | Research React | Running |
| Ashigaru 2 | Research Vue | Running |
| Ashigaru 3 | Research Angular | Completed |

Detailed flow

You: "Research the top 5 MCP servers and create a comparison table"

The Shogun writes the task to queue/shogun_to_karo.yaml and wakes the Karo. Control returns to you immediately.

The Karo breaks the task into subtasks:

All 5 Ashigaru research simultaneously. You can watch them work in real time:

Results appear in dashboard.md as they complete.

Key Features

⚡ 1. Parallel Execution

One command spawns up to 8 parallel tasks:

You: "Research 5 MCP servers"
→ 5 Ashigaru start researching simultaneously
→ Results in minutes, not hours

🔄 2. Non-Blocking Workflow

The Shogun delegates instantly and returns control to you:

You: Command → Shogun: Delegates → You: Give next command immediately
                                       ↓
                       Workers: Execute in background
                                       ↓
                       Dashboard: Shows results

No waiting for long tasks to finish.

🧠 3. Cross-Session Memory (Memory MCP)

Your AI remembers your preferences:

Session 1: Tell it "I prefer simple approaches"
            → Saved to Memory MCP

Session 2: AI loads memory on startup
            → Stops suggesting complex solutions

📡 4. Event-Driven Communication (Zero Polling)

Agents talk to each other by writing YAML files — like passing notes. No polling loops, no wasted API calls.

Karo wants to wake Ashigaru 3:

Step 1: Write the message          Step 2: Wake the agent up
┌──────────────────────┐           ┌──────────────────────────┐
│ inbox_write.sh       │           │ inbox_watcher.sh         │
│                      │           │                          │
│ Writes full message  │  file     │ Detects file change      │
│ to ashigaru3.yaml    │──change──▶│ (inotifywait, not poll)  │
│ with flock (no race) │           │                          │
└──────────────────────┘           │ Wakes agent via:         │
                                   │  1. Self-watch (skip)    │
                                   │  2. tmux send-keys       │
                                   │     (short nudge only)   │
                                   └──────────────────────────┘

Step 3: Agent reads its own inbox
┌──────────────────────────────────┐
│ Ashigaru 3 reads ashigaru3.yaml  │
│ → Finds unread messages          │
│ → Processes them                 │
│ → Marks as read                  │
└──────────────────────────────────┘

How the wake-up works:

Agent Self-Watch (ASW) Phases — Controls how aggressively the system uses tmux send-keys nudges:

Phase 2 uses the idle flag file (/tmp/shogun_idle_{agent}) to distinguish busy vs idle agents. The Stop hook creates/removes this flag at turn boundaries. This eliminates nudge interruptions during active work while still waking idle agents.

Why can't nudges be fully eliminated? Claude Code's Stop hook only fires at turn end. An idle agent (sitting at the prompt) has no turn ending, so there's no hook to trigger inbox checks. A future Notification hook with idle_prompt blocking support or a periodic timer hook could solve this.

Configure in config/settings.yaml:

asw_phase: 2   # Recommended for Claude Code setups

Or set the default directly in scripts/inbox_watcher.sh (ASW_PHASE variable). Restart inbox_watcher processes after changing.

3-Phase Escalation (v3.2) — If agent doesn't respond:

Key design choices:

• Message content is never sent through tmux — only a short "you have mail" nudge. The agent reads its own file. This eliminates character corruption and transmission hangs.
• Zero CPU while idle — inotifywait blocks on a kernel event (not a poll loop). CPU usage is 0% between messages.
• Guaranteed delivery — If the file write succeeded, the message is there. No lost messages, no retries needed.

📊 5. Agent Status Check

See which agents are busy or idle — instantly, from one command:

# Project mode: full status with task/inbox info
bash scripts/agent_status.sh

# Standalone mode: works with any tmux session
bash scripts/agent_status.sh --session mysession --lang en

Project mode output:

Agent      CLI     Pane      Task ID                                    Status     Inbox
---------- ------- --------- ------------------------------------------ ---------- -----
karo       claude  待機中    ---                                        ---        0
ashigaru1  codex   稼働中    subtask_042a_research                      assigned   0
ashigaru2  codex   待機中    subtask_042b_review                        done       0
gunshi     claude  稼働中    subtask_042c_analysis                      assigned   0

Standalone mode output (no project config needed):

Pane                           State      Agent ID
------------------------------ ---------- ----------
multiagent:agents.0            IDLE       karo
multiagent:agents.1            BUSY       ashigaru1
multiagent:agents.8            BUSY       gunshi

Detection works for both Claude Code and Codex CLI by checking CLI-specific prompt/spinner patterns in the bottom 5 lines of each tmux pane. The detection logic lives in lib/agent_status.sh — source it in your own scripts:

source lib/agent_status.sh
agent_is_busy_check "multiagent:agents.3" && echo "busy" || echo "idle"

📸 6. Screenshot Integration

VSCode's Claude Code extension lets you paste screenshots to explain issues. This CLI system provides the same capability:

# Set your screenshot folder in config/settings.yaml
screenshot:
  path: "/mnt/c/Users/YourName/Pictures/Screenshots"

# Just tell the Shogun:
You: "Check the latest screenshot"
You: "Look at the last 2 screenshots"
→ AI instantly reads and analyzes your screen captures

Windows tip: Press Win + Shift + S to take screenshots. Set the save path in settings.yaml for seamless integration.

Use cases:

• Explain UI bugs visually
• Show error messages
• Compare before/after states

📁 7. Context Management (4-Layer Architecture)

Efficient knowledge sharing through a four-layer context system:

Persistent Agent Memory (memory/MEMORY.md)

Shogun reads memory/MEMORY.md at every session start. It contains Lord's preferences, lessons learned, and cross-session knowledge — written by Shogun, read by Shogun.

┌─────────────────────────────────────────────────────────────┐
│                    Git Repositories                          │
│                                                              │
│  ┌─────────────────────┐   ┌──────────────────────────┐    │
│  │  multi-agent-shogun │   │      shogun-private        │    │
│  │       (public OSS)  │   │   (your private repo)      │    │
│  │                     │   │                            │    │
│  │ scripts/            │   │ projects/client.yaml  ←──┐ │    │
│  │ instructions/       │   │ context/my-notes.md   ←──┤ │    │
│  │ lib/                │   │ queue/shogun_to_karo.yaml │ │    │
│  │ memory/             │   │ memory/MEMORY.md      ←──┘ │    │
│  │  ├─ MEMORY.md.sample│   │ config/settings.yaml       │    │
│  │  └─ MEMORY.md  ─────┼───┼── same file, tracked here  │    │
│  │     (gitignored)    │   │                            │    │
│  └─────────────────────┘   └──────────────────────────┘    │
│         ↑ anyone can fork        ↑ your data, your repo      │
└─────────────────────────────────────────────────────────────┘

How it works: memory/MEMORY.md lives in the same working directory as the OSS repo, but is excluded from the OSS .gitignore (whitelist-based). You track it in a separate private repo using a bare git repo technique:

# One-time setup (already done by first_setup.sh)
git init --bare ~/.shogun-private.git
alias privategit='git --git-dir=$HOME/.shogun-private.git --work-tree=/path/to/multi-agent-shogun'
privategit remote add origin https://github.com/YOU/shogun-private.git

# Daily use
privategit add -f memory/MEMORY.md projects/my-client.yaml
privategit commit -m "update memory"
privategit push

The OSS .gitignore uses a whitelist approach (default: exclude everything, then explicitly allow OSS files). So private files like memory/MEMORY.md are automatically excluded without needing explicit gitignore entries — just don't add them to the whitelist.

This design enables:

• Any Ashigaru can work on any project
• Context persists across agent switches
• Clear separation of concerns
• Knowledge survives across sessions

/clear Protocol (Cost Optimization)

As agents work, their session context (Layer 4) grows, increasing API costs. /clear wipes session memory and resets costs. Layers 1–3 persist as files, so nothing is lost.

Recovery cost after /clear: ~6,800 tokens (42% improved from v1 — CLAUDE.md YAML conversion + English-only instructions reduced token cost by 70%)

1. CLAUDE.md (auto-loaded) → recognizes itself as part of the Shogun System
2. tmux display-message -t "$TMUX_PANE" -p '#{@agent_id}' → identifies its own number
3. Memory MCP read → restores the Lord's preferences (~700 tokens)
4. Task YAML read → picks up the next assignment (~800 tokens)

The key insight: designing what not to load is what drives cost savings.

Universal Context Template

All projects use the same 7-section template:

This unified format enables:

• Quick onboarding for any agent
• Consistent information management across all projects
• Easy handoff between Ashigaru workers

📱 8. Phone Notifications (ntfy)

Two-way communication between your phone and the Shogun — no SSH, no Tailscale, no server needed.

📱 You (from bed)          🏯 Shogun
    │                          │
    │  "Research React 19"     │
    ├─────────────────────────►│
    │    (ntfy message)        │  → Delegates to Karo → Ashigaru work
    │                          │
    │  "✅ cmd_042 complete"   │
    │◄─────────────────────────┤
    │    (push notification)   │

Setup:

1. Add ntfy_topic: "shogun-yourname" to config/settings.yaml
2. Install the ntfy app on your phone and subscribe to the same topic
3. shutsujin_departure.sh automatically starts the listener — no extra steps

Notification examples:

Free, no account required, no server to maintain. Uses ntfy.sh — an open-source push notification service.

⚠️ Security: Your topic name is your password. Anyone who knows it can read your notifications and send messages to your Shogun. Choose a hard-to-guess name and never share it publicly (e.g., in screenshots, blog posts, or GitHub commits).

Verify it works:

# Send a test notification to your phone
bash scripts/ntfy.sh "Test notification from Shogun 🏯"

If your phone receives the notification, you're all set. If not, check:

• config/settings.yaml has ntfy_topic set (not empty, no extra quotes)
• The ntfy app on your phone is subscribed to the exact same topic name
• Your phone has internet access and ntfy notifications are enabled

Sending commands from your phone:

1. Open the ntfy app on your phone
2. Tap your subscribed topic
3. Type a message (e.g., Research React 19 best practices) and send
4. ntfy_listener.sh receives it, writes to queue/ntfy_inbox.yaml, and wakes the Shogun
5. The Shogun reads the message and processes it through the normal Karo → Ashigaru pipeline

Any text you send becomes a command. Write it like you'd talk to the Shogun — no special syntax needed.

Manual listener start (if not using shutsujin_departure.sh):

# Start the listener in the background
nohup bash scripts/ntfy_listener.sh &>/dev/null &

# Check if it's running
pgrep -f ntfy_listener.sh

# View listener logs (stderr output)
bash scripts/ntfy_listener.sh  # Run in foreground to see logs

The listener automatically reconnects if the connection drops. shutsujin_departure.sh starts it automatically on deployment — you only need manual start if you skipped the deployment script.

Troubleshooting:

Real-world notification screenshots:

Left: Bidirectional phone ↔ Shogun communication · Right: Real-time progress report from Ashigaru

Left: Command completion notification · Right: All 8 Ashigaru completing in parallel

Note: Topic names shown in screenshots are examples. Use your own unique topic name.

SayTask Notifications

Behavioral psychology-driven motivation through your notification feed:

• Streak tracking: Consecutive completion days counted in saytask/streaks.yaml — maintaining streaks leverages loss aversion to sustain momentum
• Eat the Frog 🐸: The hardest task of the day is marked as the "Frog." Completing it triggers a special celebration notification
• Daily progress: 12/12 tasks today — visual completion feedback reinforces the Arbeitslust effect (joy of work-in-progress)

🖼️ 9. Pane Border Task Display

Each tmux pane shows the agent's current task directly on its border:

┌ ashigaru1 Sonnet+T VF requirements ──┬ ashigaru3 Opus+T API research ──────┐
│                                      │                                     │
│  Working on SayTask requirements     │  Researching REST API patterns      │
│                                      │                                     │
├ ashigaru2 Sonnet ───────────────────┼ ashigaru4 Spark DB schema design ───┤
│                                      │                                     │
│  (idle — waiting for assignment)     │  Designing database schema          │
│                                      │                                     │
└──────────────────────────────────────┴─────────────────────────────────────┘

• Working: ashigaru1 Sonnet+T VF requirements — agent name, model (with Thinking indicator), and task summary
• Idle: ashigaru2 Sonnet — model name only, no task
• Display names: Sonnet, Opus, Haiku, Codex, Spark — +T suffix = Extended Thinking enabled
• Updated automatically by the Karo when assigning or completing tasks
• Glance at all 9 panes to instantly know who's doing what

🔊 10. Shout Mode (Battle Cries)

When an Ashigaru completes a task, it shouts a personalized battle cry in the tmux pane — a visual reminder that your army is working hard.

┌ ashigaru1 (Sonnet) ──────────┬ ashigaru2 (Sonnet) ──────────┐
│                               │                               │
│  ⚔️ 足軽1号、先陣切った！     │  🔥 足軽2号、二番槍の意地！   │
│  八刃一志！                   │  八刃一志！                   │
│  ❯                            │  ❯                            │
└───────────────────────────────┴───────────────────────────────┘

How it works:

The Karo writes an echo_message field in each task YAML. After completing all work (report + inbox notification), the Ashigaru runs echo as its final action. The message stays visible above the ❯ prompt.

# In the task YAML (written by Karo)
task:
  task_id: subtask_001
  description: "Create comparison table"
  echo_message: "🔥 足軽1号、先陣を切って参る！八刃一志！"

Shout mode is the default. To disable (saves API tokens on the echo call):

./shutsujin_departure.sh --silent    # No battle cries
./shutsujin_departure.sh             # Default: shout mode (battle cries enabled)

Silent mode sets DISPLAY_MODE=silent as a tmux environment variable. The Karo checks this when writing task YAMLs and omits the echo_message field.

🗣️ SayTask — Task Management for People Who Hate Task Management

What is SayTask?

Task management for people who hate task management. Just speak to your phone.

Talk Coding, not Vibe Coding. Speak your tasks, AI organizes them. No typing, no opening apps, no friction.

• Target audience: People who installed Todoist but stopped opening it after 3 days
• Your enemy isn't other apps — it's doing nothing. The competition is inaction, not another productivity tool
• Zero UI. Zero typing. Zero app-opening. Just talk

"Your enemy isn't other apps — it's doing nothing."

How it Works

1. Install the ntfy app (free, no account needed)
2. Speak to your phone: "dentist tomorrow", "invoice due Friday"
3. AI auto-organizes → morning notification: "here's your day"

 🗣️ "Buy milk, dentist tomorrow, invoice due Friday"
       │
       ▼
 ┌──────────────────┐
 │  ntfy → Shogun   │  AI auto-categorize, parse dates, set priorities
 └────────┬─────────┘
          │
          ▼
 ┌──────────────────┐
 │   tasks.yaml     │  Structured storage (local, never leaves your machine)
 └────────┬─────────┘
          │
          ▼
 📱 Morning notification:
    "Today: 🐸 Invoice due · 🦷 Dentist 3pm · 🛒 Buy milk"

Before / After

Before (v1)	After (v2)
Raw task dump	Clean, organized daily summary

Note: Topic names shown in screenshots are examples. Use your own unique topic name.

Use Cases

• 🛏️ In bed: "Gotta submit the report tomorrow" — captured before you forget, no fumbling for a notebook
• 🚗 While driving: "Don't forget the estimate for client A" — hands-free, eyes on the road
• 💻 Mid-work: "Oh, need to buy milk" — dump it instantly and stay in flow
• 🌅 Wake up: Today's tasks already waiting in your notifications — no app to open, no inbox to check
• 🐸 Eat the Frog: AI picks your hardest task each morning — ignore it or conquer it first

FAQ

Q: How is this different from other task apps? A: You never open an app. Just speak. Zero friction. Most task apps fail because people stop opening them. SayTask removes that step entirely.

Q: Can I use SayTask without the full Shogun system? A: SayTask is a feature of Shogun. Shogun also works as a standalone multi-agent development platform — you get both capabilities in one system.

Q: What's the Frog 🐸? A: Every morning, AI picks your hardest task — the one you'd rather avoid. Tackle it first (the "Eat the Frog" method) or ignore it. Your call.

Q: Is it free? A: Everything is free and open-source. ntfy is free too. No account, no server, no subscription.

Q: Where is my data stored? A: Local YAML files on your machine. Nothing is sent to the cloud. Your tasks never leave your device.

Q: What if I say something vague like "that thing for work"? A: AI does its best to categorize and schedule it. You can always refine later — but the point is capturing the thought before it disappears.

SayTask vs cmd Pipeline

Shogun has two complementary task systems:

SayTask handles personal productivity (capture → schedule → remind). The cmd pipeline handles complex work (research, code, multi-step tasks). Both share streak tracking — completing either type of task counts toward your daily streak.

Model Settings

Thinking control: Set thinking: true/false per agent in config/settings.yaml. When thinking: false, the agent starts with MAX_THINKING_TOKENS=0 to disable Extended Thinking. Pane borders show +T suffix when Thinking is enabled (e.g., Sonnet+T, Opus+T).

Live model switching: Use /shogun-model-switch to change any agent's CLI type, model, or Thinking setting without restarting the entire system. See the Skills section for details.

The system routes work by cognitive complexity at two levels: Agent routing (Ashigaru for L1–L3, Gunshi for L4–L6) and Model routing within Ashigaru via capability_tiers (see Dynamic Model Routing below).

Bloom's Taxonomy → Agent Routing

Tasks are classified using Bloom's Taxonomy and routed to the appropriate agent, not model:

The Karo assigns each subtask a Bloom level and routes it to the appropriate agent. L1–L3 tasks go to Ashigaru for parallel execution; L4–L6 tasks go to the Gunshi for deeper analysis. Simple L4 tasks (e.g., small code review) may still go to Ashigaru when the Karo judges it appropriate.

Task Dependencies (blockedBy)

Tasks can declare dependencies on other tasks using blockedBy:

# queue/tasks/ashigaru2.yaml
task:
  task_id: subtask_010b
  blockedBy: ["subtask_010a"]  # Waits for ashigaru1's task to complete
  description: "Integrate the API client built by subtask_010a"

When a blocking task completes, the Karo automatically unblocks dependent tasks and assigns them to available Ashigaru. This prevents idle waiting and enables efficient pipelining of dependent work.

Dynamic Model Routing (capability_tiers)

Beyond agent-level routing, you can configure model-level routing within the Ashigaru tier. Define a capability_tiers table in config/settings.yaml mapping each model to its maximum Bloom level:

capability_tiers:
  gpt-5.3-codex-spark:
    max_bloom: 3       # L1–L3 only: fast, high-volume tasks
    cost_group: chatgpt_pro
  gpt-5.3-codex:
    max_bloom: 4       # L1–L4: + analysis and debugging
    cost_group: chatgpt_pro
  claude-sonnet-4-6:
    max_bloom: 5       # L1–L5: + design evaluation
    cost_group: claude_max
  claude-opus-4-6:
    max_bloom: 6       # L1–L6: + novel architecture, strategy
    cost_group: claude_max

The cost_group field links each model to your subscription plan, enabling the system to avoid routing tasks to models your plan doesn't cover.

Two built-in skills help you configure this:

Run /shogun-bloom-config after setup to generate your optimal capability_tiers configuration.

Philosophy

"Don't execute tasks mindlessly. Always keep 'fastest × best output' in mind."

The Shogun System is built on five core principles:

These principles are documented in detail: docs/philosophy.md

Design Philosophy

Why a hierarchy (Shogun → Karo → Ashigaru)?

1. Instant response: The Shogun delegates immediately, returning control to you
2. Parallel execution: The Karo distributes to multiple Ashigaru simultaneously
3. Single responsibility: Each role is clearly separated — no confusion
4. Scalability: Adding more Ashigaru doesn't break the structure
5. Fault isolation: One Ashigaru failing doesn't affect the others
6. Unified reporting: Only the Shogun communicates with you, keeping information organized

Why Mailbox System?

Why use files instead of direct messaging between agents?

Agent Identification (@agent_id)

Each pane has a @agent_id tmux user option (e.g., karo, ashigaru1). While pane_index can shift when panes are rearranged, @agent_id is set at startup by shutsujin_departure.sh and never changes.

Agent self-identification:

tmux display-message -t "$TMUX_PANE" -p '#{@agent_id}'

The -t "$TMUX_PANE" is required. Omitting it returns the active pane's value (whichever pane you're focused on), causing misidentification.

Model names are stored as @model_name and current task summaries as @current_task — both displayed in the pane-border-format. Even if Claude Code overwrites the pane title, these user options persist.

Why only the Karo updates dashboard.md

1. Single writer: Prevents conflicts by limiting updates to one agent
2. Information aggregation: The Karo receives all Ashigaru reports, so it has the full picture
3. Consistency: All updates pass through a single quality gate
4. No interruptions: If the Shogun updated it, it could interrupt the Lord's input

Skills

No skills are included out of the box. Skills emerge organically during operation — you approve candidates from dashboard.md as they're discovered.

Invoke skills with /skill-name. Just tell the Shogun: "run /skill-name".

Included Skills (committed to repo)

Skills ship with the repository in skills/. They are domain-agnostic utilities useful for any user:

These help you configure and operate the system. Personal workflow skills grow organically through the bottom-up discovery process.

Skill Philosophy

1. Personal skills are not committed to the repo

Skills in .claude/commands/ are excluded from version control by design:

• Every user's workflow is different
• Rather than imposing generic skills, each user grows their own skill set

2. How skills are discovered

Ashigaru notices a pattern during work
    ↓
Appears in dashboard.md under "Skill Candidates"
    ↓
You (the Lord) review the proposal
    ↓
If approved, instruct the Karo to create the skill

Skills are user-driven. Automatic creation would lead to unmanageable bloat — only keep what you find genuinely useful.

MCP Setup Guide

MCP (Model Context Protocol) servers extend Claude's capabilities. Here's how to set them up:

What is MCP?

MCP servers give Claude access to external tools:

• Notion MCP → Read and write Notion pages
• GitHub MCP → Create PRs, manage issues
• Memory MCP → Persist memory across sessions

Installing MCP Servers

Add MCP servers with these commands:

# 1. Notion - Connect to your Notion workspace
claude mcp add notion -e NOTION_TOKEN=your_token_here -- npx -y @notionhq/notion-mcp-server

# 2. Playwright - Browser automation
claude mcp add playwright -- npx @playwright/mcp@latest
# Note: Run `npx playwright install chromium` first

# 3. GitHub - Repository operations
claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=your_pat_here -- npx -y @modelcontextprotocol/server-github

# 4. Sequential Thinking - Step-by-step reasoning for complex problems
claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking

# 5. Memory - Cross-session long-term memory (recommended!)
# ✅ Auto-configured by first_setup.sh
# To reconfigure manually:
claude mcp add memory -e MEMORY_FILE_PATH="$PWD/memory/shogun_memory.jsonl" -- npx -y @modelcontextprotocol/server-memory

Verify installation

claude mcp list

All servers should show "Connected" status.

Real-World Use Cases

This system manages all white-collar tasks, not just code. Projects can live anywhere on your filesystem.

Example 1: Research sprint

You: "Research the top 5 AI coding assistants and compare them"

What happens:
1. Shogun delegates to Karo
2. Karo assigns:
   - Ashigaru 1: Research GitHub Copilot
   - Ashigaru 2: Research Cursor
   - Ashigaru 3: Research Claude Code
   - Ashigaru 4: Research Codeium
   - Ashigaru 5: Research Amazon CodeWhisperer
3. All 5 research simultaneously
4. Results compiled in dashboard.md

Example 2: PoC preparation

You: "Prepare a PoC for the project on this Notion page: [URL]"

What happens:
1. Karo fetches Notion content via MCP
2. Ashigaru 2: Lists items to verify
3. Ashigaru 3: Investigates technical feasibility
4. Ashigaru 4: Drafts a PoC plan
5. All results compiled in dashboard.md — meeting prep done

Configuration

Language

# config/settings.yaml
language: ja   # Samurai Japanese only
language: en   # Samurai Japanese + English translation

Screenshot integration

# config/settings.yaml
screenshot:
  path: "/mnt/c/Users/YourName/Pictures/Screenshots"

Tell the Shogun "check the latest screenshot" and it reads your screen captures for visual context. (Win+Shift+S on Windows.)

ntfy (Phone Notifications)

# config/settings.yaml
ntfy_topic: "shogun-yourname"

Subscribe to the same topic in the ntfy app on your phone. The listener starts automatically with shutsujin_departure.sh.

ntfy Authentication (Self-Hosted Servers)

The public ntfy.sh instance requires no authentication — the setup above is all you need.

If you run a self-hosted ntfy server with access control enabled, configure authentication:

# 1. Copy the sample config
cp config/ntfy_auth.env.sample config/ntfy_auth.env

# 2. Edit with your credentials (choose one method)

Priority: Token > Basic > None. If neither is set, no auth headers are sent (backward compatible).

config/ntfy_auth.env is excluded from git. See config/ntfy_auth.env.sample for details.

Advanced

┌─────────────────────────────────────────────────────────────────────┐
│                    First-Time Setup (run once)                       │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  install.bat (Windows)                                              │
│      │                                                              │
│      ├── Check/guide WSL2 installation                              │
│      └── Check/guide Ubuntu installation                            │
│                                                                     │
│  first_setup.sh (run manually in Ubuntu/WSL)                        │
│      │                                                              │
│      ├── Check/install tmux                                         │
│      ├── Check/install Node.js v20+ (via nvm)                      │
│      ├── Check/install Claude Code CLI (native version)             │
│      │       ※ Proposes migration if npm version detected           │
│      └── Configure Memory MCP server                                │
│                                                                     │
├─────────────────────────────────────────────────────────────────────┤
│                    Daily Startup (run every day)                     │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  shutsujin_departure.sh                                             │
│      │                                                              │
│      ├──▶ Create tmux sessions                                      │
│      │         • "shogun" session (1 pane)                          │
│      │         • "multiagent" session (9 panes, 3x3 grid)          │
│      │                                                              │
│      ├──▶ Reset queue files and dashboard                           │
│      │                                                              │
│      └──▶ Launch Claude Code on all agents                          │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

# Default: Full startup (tmux sessions + Claude Code launch)
./shutsujin_departure.sh

# Session setup only (no Claude Code launch)
./shutsujin_departure.sh -s
./shutsujin_departure.sh --setup-only

# Clean task queues (preserves command history)
./shutsujin_departure.sh -c
./shutsujin_departure.sh --clean

# Battle formation: All Ashigaru on Opus (max capability, higher cost)
./shutsujin_departure.sh -k
./shutsujin_departure.sh --kessen

# Silent mode: Disable battle cries (saves API tokens on echo calls)
./shutsujin_departure.sh -S
./shutsujin_departure.sh --silent

# Full startup + open Windows Terminal tabs
./shutsujin_departure.sh -t
./shutsujin_departure.sh --terminal

# Shogun relay-only mode: Disable Shogun's thinking (cost savings)
./shutsujin_departure.sh --shogun-no-thinking

# Show help
./shutsujin_departure.sh -h
./shutsujin_departure.sh --help

./shutsujin_departure.sh          # Launch everything
tmux attach-session -t shogun     # Connect and give commands

./shutsujin_departure.sh -s       # Create sessions only

# Manually launch Claude Code on specific agents
tmux send-keys -t shogun:0 'claude --dangerously-skip-permissions' Enter
tmux send-keys -t multiagent:0.0 'claude --dangerously-skip-permissions' Enter

# Kill existing sessions
tmux kill-session -t shogun
tmux kill-session -t multiagent

# Fresh start
./shutsujin_departure.sh

alias csst='cd /mnt/c/tools/multi-agent-shogun && ./shutsujin_departure.sh'
alias css='tmux attach-session -t shogun'      # Connect to Shogun
alias csm='tmux attach-session -t multiagent'  # Connect to Karo + Ashigaru

File Structure

multi-agent-shogun/
│
│  ┌──────────────── Setup Scripts ────────────────────┐
├── install.bat               # Windows: First-time setup
├── first_setup.sh            # Ubuntu/Mac: First-time setup
├── shutsujin_departure.sh    # Daily deployment (auto-loads instructions)
│  └──────────────────────────────────────────────────┘
│
├── instructions/             # Agent behavior definitions
│   ├── shogun.md             # Shogun instructions
│   ├── karo.md               # Karo instructions
│   ├── ashigaru.md           # Ashigaru instructions
│   ├── gunshi.md             # Gunshi (strategist) instructions
│   └── cli_specific/         # CLI-specific tool descriptions
│       ├── claude_tools.md   # Claude Code tools & features
│       └── copilot_tools.md  # GitHub Copilot CLI tools & features
│
├── lib/
│   ├── agent_status.sh       # Shared busy/idle detection (Claude Code + Codex)
│   ├── cli_adapter.sh        # Multi-CLI adapter (Claude/Codex/Copilot/Kimi)
│   └── ntfy_auth.sh          # ntfy authentication helper
│
├── scripts/                  # Utility scripts
│   ├── agent_status.sh       # Show busy/idle status of all agents
│   ├── inbox_write.sh        # Write messages to agent inbox
│   ├── inbox_watcher.sh      # Watch inbox changes via inotifywait
│   ├── switch_cli.sh         # Live CLI/model switching (/exit → relaunch)
│   ├── ntfy.sh               # Send push notifications to phone
│   └── ntfy_listener.sh      # Stream incoming messages from phone
│
├── config/
│   ├── settings.yaml         # Language, ntfy, and other settings
│   ├── ntfy_auth.env.sample  # ntfy authentication template (self-hosted)
│   └── projects.yaml         # Project registry
│
├── projects/                 # Project details (excluded from git, contains confidential info)
│   └── <project_id>.yaml    # Full info per project (clients, tasks, Notion links, etc.)
│
├── queue/                    # Communication files
│   ├── shogun_to_karo.yaml   # Shogun → Karo commands
│   ├── ntfy_inbox.yaml       # Incoming messages from phone (ntfy)
│   ├── inbox/                # Per-agent inbox files
│   │   ├── shogun.yaml       # Messages to Shogun
│   │   ├── karo.yaml         # Messages to Karo
│   │   └── ashigaru{1-8}.yaml # Messages to each Ashigaru
│   ├── tasks/                # Per-worker task files
│   └── reports/              # Worker reports
│
├── saytask/                  # Behavioral psychology-driven motivation
│   └── streaks.yaml          # Streak tracking and daily progress
│
├── templates/                # Report and context templates
│   ├── integ_base.md         # Integration: base template
│   ├── integ_fact.md         # Integration: fact-finding
│   ├── integ_proposal.md     # Integration: proposal
│   ├── integ_code.md         # Integration: code review
│   ├── integ_analysis.md     # Integration: analysis
│   └── context_template.md   # Universal 7-section project context
│
├── skills/                   # Reusable skills (committed to repo)
│   ├── skill-creator/        # Skill creation template
│   ├── shogun-agent-status/  # Agent status display
│   ├── shogun-model-list/    # Model capability reference
│   ├── shogun-bloom-config/  # Bloom tier configurator
│   ├── shogun-model-switch/  # Live CLI/model switching
│   └── shogun-readme-sync/   # README sync
│
├── memory/                   # Memory MCP persistent storage
├── dashboard.md              # Real-time status board
└── CLAUDE.md                 # System instructions (auto-loaded)

Project Management

This system manages not just its own development, but all white-collar tasks. Project folders can be located outside this repository.

How it works

config/projects.yaml          # Project list (ID, name, path, status only)
projects/<project_id>.yaml    # Full details for each project

• config/projects.yaml: A summary list of what projects exist
• projects/<id>.yaml: Complete details (client info, contracts, tasks, related files, Notion pages, etc.)
• Project files (source code, documents, etc.) live in the external folder specified by path
• projects/ is excluded from git (contains confidential client information)

Example

# config/projects.yaml
projects:
  - id: client_x
    name: "Client X Consulting"
    path: "/mnt/c/Consulting/client_x"
    status: active

# projects/client_x.yaml
id: client_x
client:
  name: "Client X"
  company: "X Corporation"
contract:
  fee: "monthly"
current_tasks:
  - id: task_001
    name: "System Architecture Review"
    status: in_progress

This separation lets the Shogun System coordinate across multiple external projects while keeping project details out of version control.

Troubleshooting

# Re-run first_setup.sh
./first_setup.sh

# If npm version is detected:
# ⚠️ npm version of Claude Code CLI detected (officially deprecated)
# Install native version? [Y/n]:

# After selecting Y, uninstall npm version:
npm uninstall -g @anthropic-ai/claude-code

ToolSearch("select:mcp__memory__read_graph")
mcp__memory__read_graph()

tmux attach-session -t multiagent
# Ctrl+B then 0-8 to switch panes

# Method 1: Run claude directly in the pane
claude --model opus --dangerously-skip-permissions

# Method 2: Karo force-restarts via respawn-pane (also fixes nesting)
tmux respawn-pane -t shogun:0.0 -k 'claude --model opus --dangerously-skip-permissions'

tmux Quick Reference

Mouse Support

first_setup.sh automatically configures set -g mouse on in ~/.tmux.conf, enabling intuitive mouse control:

Even if you're not comfortable with keyboard shortcuts, you can switch, scroll, and resize panes using just the mouse.

What's New in v3.5 — Dynamic Model Routing

Right model for the right task — without restarting any agent. Sonnet 4.6 closes the gap with Opus to just 1.2pp on SWE-bench (79.6% vs 80.8%), making per-task model routing practical and cost-effective for the first time.

• Bloom Dynamic Model Routing — capability_tiers in config/settings.yaml maps each model to its Bloom ceiling. L1–L3 → Spark (1000+ tok/s), L4 → Sonnet 4.6, L5 → Sonnet 4.6 + extended thinking, L6 → Opus (genuinely novel design only). Routing happens without agent restarts — the system finds the right idle agent by model capability
• Sonnet 4.6 as the new standard — SWE-bench 79.6%, only 1.2pp below Opus 4.6. Gunshi downgraded Opus → Sonnet 4.6. All Ashigaru default to Sonnet 4.6. One YAML line change, no restarts required
• /shogun-model-list skill — Complete reference table: all CLI tools × models × subscriptions × Bloom max level. Updated for Sonnet 4.6 and Spark positioning
• /shogun-bloom-config skill — Interactive configurator: answer 2 questions about your subscriptions → get ready-to-paste capability_tiers YAML

What's New in v3.3.2 — GPT-5.3-Codex-Spark Support

New model, same YAML. Add model: gpt-5.3-codex-spark to any Codex agent in settings.yaml.

• Codex --model flag support — build_cli_command() now passes settings.yaml model config to the Codex CLI via --model. Supports gpt-5.3-codex-spark and any future Codex models
• Separate rate limit — Spark runs on its own rate limit quota, independent of GPT-5.3-Codex. Run both models in parallel across different Ashigaru to double your effective throughput
• Startup display — shutsujin_departure.sh now shows the actual model name (e.g., codex/gpt-5.3-codex-spark) instead of the generic effort level

What's New in v3.0 — Multi-CLI

Shogun is no longer Claude-only. Mix and match 4 AI coding CLIs in a single army.

• Multi-CLI as first-class architecture — lib/cli_adapter.sh dynamically selects CLI per agent. Change one line in settings.yaml to swap any worker between Claude Code, Codex, Copilot, or Kimi
• OpenAI Codex CLI integration — GPT-5.3-codex with --dangerously-bypass-approvals-and-sandbox for true autonomous execution. --no-alt-screen makes agent activity visible in tmux
• CLI bypass flag discovery — --full-auto is NOT fully automatic (it's -a on-request). Documented the correct flags for all 4 CLIs
• Hybrid architecture — Command layer (Shogun + Karo) stays on Claude Code for Memory MCP and mailbox integration. Worker layer (Ashigaru) is CLI-agnostic
• Community-contributed CLI adapters — Thanks to @yuto-ts (cli_adapter.sh), @circlemouth (Codex support), @koba6316 (task routing)

Sponsors

This project is funded by sponsors. Your support keeps it free and actively maintained.

Contributing

Issues and pull requests are welcome.

• Bug reports: Open an issue with reproduction steps
• Feature ideas: Open a discussion first
• Skills: Skills are personal by design and not included in this repo

Credits

Based on Claude-Code-Communication by Akira-Papa.

License

MIT