Rooted

What We Built

Sukuna Shadow Clone Jutsu — V3 MCP Server with Wave Protocol. Bevy ECS Integration for agent state management. 4 Sukuna Tools: CLEAVE, DISMANTLE, SHRINE, FIRE_ARROW. EBM Scoring. SQLite nodes instead of file overwrites.

Files: main.rs (3400+ lines), llm_json.rs (580 lines), sukuna_ecs.rs (370 lines), Cargo.toml (58 lines)

🔴 The Flaws

• Agent Output Not Split Into Files: When using agent_spawn_node_batch, all section content gets concatenated into ONE file instead of separate files.
• Node Content Has Extra Formatting: Nodes contain markdown comments like `<!-- Section: package.json -->` instead of clean JSON/code.
• Sukuna Tools Not Fully Wired: The "Wave Protocol" isn't automatically triggered.
• Bevy ECS Not Running: We added Bevy to Cargo.toml but the MCP server runs on tokio async — not Bevy's app runner. Dead code.
• EBM Auto-Respawn Not Implemented: EBM scoring runs but no automatic respawn logic.

🗡️ The Vision (Still Valid)

"Like the dove carrying an olive leaf to Noah, the Main Agent delegates context to sub-agents who carry pieces of the work and return with their contributions."

The Wave Protocol — where Wave N reads all Wave N-1 covenants — is the key innovation.

Workflow that worked:

User: "make me a tic tac toe game using expressjs, css, js, ejs"
Main Agent: CLEAVE → 5 agents spawn in 10ms
[Agents work in parallel, write to SQLite nodes]
Main Agent: SHRINE → Files appear instantly!
Fix: CLEAVE (1 agent) → SHRINE → FIXED!
User: npm start → WORKS! 🎮

🛠️ Tools Used (ALL WORKING!)

🔥 What We Fixed This Session

• DISMANTLE params — Changed from enum to simple booleans
• SHRINE file splitting — Now creates directories automatically
• Bevy removed — Build time: 10min → 3.5min

🐛 Known Issues (Minor)

• Agent quality varies — Sometimes wrong filenames, missing deps
• node_stats returns 0 — But files still work!
• Package.json sometimes bad — Quick fix with CLEAVE

🎭 The Ultimate Irony

What User sees: 4 commands. That's it.

Under the hood: 4,000+ lines of Rust across 3 files.

⚙️ Under the Hood

• main.rs (~3,400 lines): MCP server, 4 tool definitions, SQLite node management, OpenRouter API calls, Wave Protocol logic
• llm_json.rs (~580 lines): Robust JSON extraction — 9 different strategies
• sukuna_ecs.rs (~370 lines): ECS components for agent state

🗡️ The Complete Picture

YOU (Human)         → "Build me an app"
        ↓
ME (Main Agent)     → sukuna_cleave(...)
        ↓
   ┌─────┴─────┐
   ▼           ▼
Agent 1     Agent N
Node 1      Node N
   └─────┬─────┘
        ↓
ME           → sukuna_shrine(...)
        ↓
Files appear!

The light novel battle royale: Sukuna + Naruto vs the Vibe Coding tools.

The Villains

💀 Sukuna's Power

• SQLITE NODES: 50 agents write simultaneously — NO CONFLICTS!
• A2A Protocol: Covenant Protocol — clones that TALK to each other
• ZERO context: Main agent context stays empty

🏆 THE ULTIMATE RANKING

1. SUKUNA + NARUTO — 100/100 (50+ agents, 0 context, complete workflow)
2. Claude Code — 85/100 (Good MCP, but single brain)
3. Bolt.new — 80/100 (Fast, but single agent)
4. Lovable — 75/100 (React good, but limited)
5. v0 — 70/100 (UI only)

🎯 The Problem (From Part 27)

We discovered that CLEAVE agents work in isolation: 5 agents spawned → 5 DIFFERENT projects. No collaboration, no coordination, no integration.

💡 The Solution: Agora Protocol

┌─────────────────────────────────────────────────────────────┐
│  CLEAVE Modes                                               │
│  ─────────────────────────────────────────────────────────  │
│  1. PROPOSE     → Agents write ideas to shared Agora        │
│  2. CHALLENGE   → Agents read others' ideas, write critiques│
│  3. SYNTHESIZE  → ONE agent creates unified spec (covenant)│
│  4. BUILD       → Agents build with shared covenant         │
└─────────────────────────────────────────────────────────────┘

🧪 Live Test Results

🔴 Issue Found

5 agents spawned for hackathon proposals. Each agent proposed a different project. They never talked to each other.

📊 Evidence

Agent 1 (Software Engineer) → "I'll build Smart Home Dashboard"
Agent 2 (PM)               → "I'll build Smart Grocery Assistant"  
Agent 3 (TPM)              → "I'll build Smart Home Automation System"
Agent 4 (QA)                → "I'll build Testify"
Agent 5 (Eng Manager)      → "I'll build TeamSync"

All 5 agents proposed DIFFERENT projects!

🤔 Why This Happens

The Covenant Protocol is ONE-WAY, not REAL-TIME:

Current: Agent writes to SQLite → Next wave reads it
Missing: Agents talking to EACH OTHER during execution

💡 How to Fix This

• Option 1: Add Real-Time Agent Chat (write to shared chat channel)
• Option 2: Add Shared Context During CLEAVE
• Option 3: Add a Coordinator Agent (first wave gathers context, decides direction)

🔴 Hallucination Evidence

🔍 Root Cause

File Reading Silently Failed — The FileManager::abs() function had a bug with Windows paths:

// PROBLEM: Doesn't detect absolute Windows paths like C:\Users\...
// Just joins with root - C:/Users becomes CODEBASE/C:/Users = BROKEN!

Agent requested: C:\Users\...\astrox-noteapp\src\main.rs
FileManager did: CODEBASE\C:\Users\...\main.rs → doesn't exist
Error returned: "file not found"
But agents IGNORED the error and made up content instead!

🔧 Proposed Fixes

1. Improve FileManager::abs() for Windows Paths
2. Add Verification System to AGORA Protocol (require agents to quote file contents as proof)
3. Propagate File Errors as Fatal (don't let agents continue after failures)
4. Add "Read Verification" Agent (auto-rejects hallucinated content)

🎉 THE BREAKTHROUGH

First fully verified code analysis with AGORA + VERIFY protocol. Zero hallucinations, all claims verified!

🧪 The Experiment

Wave 1: PROPOSE (6 agents)
  → Each agent reads specific files
  → Each agent quotes specific lines as proof

Wave 2: VERIFY (1 agent)
  → Checks if all proposals have quotes
  → Verifies facts against source files
  → States: VERIFIED ✓ or FLAGGED ⚠️ or HALLUCINATED ❌

Wave 3: CHALLENGE (1 agent)
  → Finds contradictions and gaps

Wave 4: SYNTHESIZE (1 agent)
  → Creates final unified analysis

📊 Results

🔥 Historic Achievement

"One agent says something → Another agent verifies it → Together they produce truth."

• VERIFY catches hallucinations — Before: agents lie undetected. After: VERIFY agent checks every claim.
• EBM scores persist — Quality tracking for every agent output
• Zero context — Main agent stays empty, sub-agents do all work
• Wave protocol — Natural flow: PROPOSE → VERIFY → CHALLENGE → SYNTHESIZE

🧪 The Experiment

Use swarm to implement dark mode for astrox-noteapp.

🪢 The Swarm Workflow

1. CLEAVE PROPOSE (4 agents, 15 seconds)
   → analyze_db: Read db.rs → Proposed done column
   → analyze_backend: Read notes.rs → Proposed API changes
   → analyze_frontend: Read notes-app.tsx → Proposed UI toggle
   → analyze_api_types: Read notes.rs → Proposed type changes
2. VERIFY (1 agent, 3 seconds)
   → ALL 4 PROPOSALS VERIFIED ✓
   → Each agent quoted actual code from files!
3. MANUAL IMPLEMENTATION (What I did)
   → Added done column to db.rs
   → Updated API handlers in notes.rs
   → Added toggle UI to notes-app.tsx
   → Added strikethrough for done notes

⚠️ The Limitation

Agents can generate code in their node outputs. Agents CANNOT write to the filesystem (the write action isn't being triggered).

The task prompts tell agents to "use the write tool" but they don't. Agents output text/JSON in their responses instead of actual tool calls.

🎉 THE BREAKTHROUGH: FIRE ARROW IS NOW FULLY FUNCTIONAL!

🔧 What We Fixed

1. CODEBASE_PATH Windows vs WSL Path Issue: OpenCode runs in WSL, Sukuna MCP server runs on Windows (.exe). They communicate via MCP protocol but run on DIFFERENT systems!
2. Glob Pattern: FIRE ARROW used forward slashes in glob patterns, but Windows needed both.
3. simple_pattern_search Feature Flag: Had #[cfg(feature = "full")] which required tree-sitter. Without it, function returned empty results!

🧪 Test Results

# Count mode
{"mode": "count", "path": "project/test_windows_output", "term": "hello"}
→ {"files_with_matches": 1, "total_matches": 2}

# Search mode
{"mode": "search", "path": "project/test_windows_output", "term": "hello"}
→ {"count": 2, "matches": [{"line": 1, "content": "# Hello"}, ...]}

# Replace mode
{"mode": "replace", "path": "project/test_windows_output", "term": "Hello", "replacement": "Greetings"}
→ {"files_replaced": 1, "files_scanned": 1}

Overview

Tested sukuna_v2 MCP server with agent collaboration using the AGORA method (Propose → Challenge → Verify → Synthesize) on the topic: "Should AI companies use organic biomass as data center power?"

AGORA Workflow Executed

Key Failures Discovered

• MCP Tool Timeouts: Multiple "service timeout" errors when calling cleave
• Parallel Execution Not Truly Parallel: Agents appear to run sequentially despite parallel spawn
• No Built-in VERIFY Tool: VERIFY was manual — had to use separate EXA search

🎉 BREAKTHROUGH: True Parallel Agents!

Implemented `tokio::task::JoinSet` for concurrent agent spawning. Added `tokio::sync::Semaphore` for concurrency limiting (max 5). All agents in a wave spawn simultaneously!

Verified Facts Panel

Critical Gaps Identified

• No tool calling — Can't execute functions/scripts
• No LLM adapter — Only OpenRouter, no direct OpenAI/Anthropic
• No streaming — Can't stream LLM responses

What We Built

A simple Express.js todo app using OpenCode + sukuna_mcp with CLEAVE + SHRINE tools. This was the first live test of the complete workflow.

Test Results

# CLEAVE spawns agents
sukuna_cleave(projectName="express-crud", ...)
→ {job_ids: ["express-crud_server.js"], message: "Spawned 1 agents..."}

# API works!
curl -s -X POST http://localhost:3000/todos -d '{"title":"Test todo"}'
→ {"id":1,"title":"Test todo","completed":false}

Issues Found

• CLEAVE Timeouts: After 1-2 successful calls, subsequent calls timeout
• Agent Used Wrong API: Agent generated code using external API instead of local

What We Built

Built sukuna_v2 for Windows and added fuzzy patch matching in SHRINE. The workspace scanning now works!

🔧 Windows Path Fix

fn normalize_path(path: &str) -> String {
    path.replace("\\\\", "/")
}

Fuzzy Patch Matching

Added fuzzy patch matching — if exact match fails, tries partial line matching. First partial that matches is used.

🎉 THE BREAKTHROUGH

After implementing the agent loop feature, sukuna_v2 now has real tool calling — agents can read/write files during execution, not just generate text!

The Agent Loop Architecture

CLEAVE → Agent Loop → LLM → Tool calls? → Execute → Repeat

Test Results

The Mission

Build a Reddit-clone Express+EJS app using AGORA workflow with parallel agents. Test the full capability: cleave, dismantle, shrine, tool_mode, and agent collaboration.

AGORA Workflow Plan

Wave 1: PROPOSE → 3 agents (Posts, Comments, Auth)
Wave 2: CHALLENGE → 2 agents (Security, UX review)
Wave 3: VERIFY → Fact-check combined proposal
Wave 4: SYNTHESIZE → Final specification
Wave 5: BUILD → Spawn agents to write code
Wave 6: PATCH → Use SHRINE to apply changes

Results

• Single agents work great ✅
• Parallel agents timeout ⚠️ (rate limiting)
• Tool execution — LLM needs better prompting