ToonGine — CAOS Cognitive Agent OS

One npm install. 25 AI agents. Full venture OS.

npx toongine init    # Deploy agents, scaffold .toon/, build graphs

---

Architecture

npx toongine init
       │
       ▼
┌──────────────────────────────────────────────────┐
│                VPS (Hermes Agent)                 │
│                                                   │
│  POST :4201/register → creates project            │
│  /root/.toon/projects/<name>/                     │
│    ├── agents/         25 agents                  │
│    ├── metrics.json    token burn, cost, sessions │
│    ├── config.json     project config + memory    │
│    └── graphs/         codegraph, graphify        │
│                                                   │
│  Pipeline (every 5 min):                          │
│    state.db → metrics.json → GitHub .toon/        │
│                                                   │
│  API (:4201):                                     │
│    /register  /projects  /metrics  /agents        │
│    /health    /metrics/all                        │
└──────────────────────────────────────────────────┘
       │
       ▼
┌──────────────────────────────────────────────────┐
│              Vercel Dashboard (yvon.in)            │
│                                                   │
│  Agents tab → Token Burn / Health / Memory        │
│  All data from VPS API — real metrics             │
└──────────────────────────────────────────────────┘

---

25 Agents — 10 Departments

Department	Agents	Role
CEO	marcus	Direction + Orchestration
COO	diana	Operations + Sprints
Command	board	Governance + Constitutional Authority
Finance	felix	Financial Intelligence
Legal	comply, docs, guard	Compliance, Docs, Security
Marketing	kai, lena, rio, nate, atlas, pixel, aria	Revenue + Content (7 agents)
Psychology	kahneman	Behavioral Intelligence + Bias Audit
Research	depth, synth, vette	Deep Research + Synthesis + Validation
Sense	forge, radar, scout	Discovery + Market Intel + Feasibility
Technical	dev, raj, mia, quinn	Architecture, Backend, Frontend, QA

---

Agent File Structure

Every agent carries:

File	Purpose
AGENT.md	Identity — who, what, personality
MEMORY.md	Learnings across sessions (fills during work)
SESSION.md	Current task state (fills during active work)
SKILLS.md	Load triggers — when to load what skill
TOOLS.md	Available tools
manifest.toon	Machine-readable spec
skills/	Custom skills + Marketplace skills + OS protocols

---

Governance Cycle (Board)

Every strategic decision flows through 4 gates:

Pre-Mortem → Decision Critic → Fiduciary Guard → Strategic Veto
     │              │                │                  │
  "What could    "Stress-test    "Can we afford    "Should we do
   go wrong?"     the logic"      this?"            this?"

Board skills:

• Custom: constitution-enforcement, fiduciary-guard, risk-assessment-matrix, strategic-veto
• Marketplace: decision-critic, postmortem-writing, pre-mortem
• OS: reflection-protocol, triple-pass-protocol

---

Pipeline

toongine-pipeline.py (cron every 5 min):

1. Reads Hermes state.db — sessions, tokens, costs, agent attribution
2. Matches sessions to projects via working directory
3. Writes per-project metrics.json (tokens, cost, hourly burn, per-agent, per-provider)
4. Syncs agent memory health stats to config.json
5. Pushes .toon/ to GitHub for visibility

Zero Supabase. Zero fake data.

---

API Endpoints (VPS :4201)

Endpoint	Returns
POST /register	Create project — deploys 25 agents
GET /projects	List all registered projects
GET /metrics?project=<name>	Per-project token burn, cost, agents
GET /metrics/all	Aggregate across all projects
GET /agents?project=<name>	Agent memory health per project
GET /health	Server status

---

Per-Project Structure

After npx toongine init, the VPS creates:

/root/.toon/projects/<name>/
├── agents/              25 agents with MEMORY.md
│   ├── CEO/marcus/
│   ├── COO/diana/
│   ├── Command/board/
│   ├── Finance/felix/
│   ├── Legal/{comply,docs,guard}/
│   ├── Marketing/{kai,lena,rio,nate,atlas,pixel,aria}/
│   ├── Psychology/Daniel_Kahneman/
│   ├── Research/{depth,synth,vette}/
│   ├── Sense/{forge,radar,scout}/
│   └── Technical/{dev,raj,mia,quinn}/
├── config.json          repo_id, version, agent memory health
├── metrics.json         token burn, cost, sessions, hourly, per-agent
├── graph/ codegraph/ graphify/ code-review-graph/
├── hermes/              bridge config
└── state/               runtime state

---

CAOS Pipelines (yvon-engine/.toon/hermes/caos/)

Module	Purpose
pipeline_router.py	Classifies intent: NEW_PROJECT / NEW_FEATURE / NEW_IDEA
research_phase.py	6-agent research gate with 5-dimension scoring
agent_protocol.py	4 checkpoints (CP1-CP4) — every agent self-questions
caos_loop.py	Marcus → Diana → Dev/Raj/Mia → Quinn → Council
caos_executor.py	Executes agent tasks with discipline gates
caos_verifier.py	Post-execution verification
discipline_gate.py	6 discipline gates before output ships
council.py	Cross-agent deliberation and conflict resolution
self_counter.py	Agent self-critique and correction loop
memory_store.py	Persistent agent memory operations
agent_registry.py	Agent manifest, capabilities, department map
project_registry.py	Multi-project isolation and routing

---

Commands

npx toongine init              # Deploy agents + scaffold project
npx toongine compile           # Build TOON compression corpus
npx toongine graphify:build    # Build AST knowledge graph
npx toongine codegraph:build   # Build import dependency graph
npx toongine project list      # List registered projects
npx toongine project switch    # Switch active project
npx toongine dashboard         # Standalone dashboard (port 4200)

---

Stack

• Agent runtime: Hermes Agent on VPS (DeepSeek v4)
• Compression: TOON v4 — 84.5% token savings
• Dashboard: Next.js 15 on Vercel
• Graphs: code-review-graph (semantic), graphify (AST), codegraph (imports)
• Pipeline: Python cron, zero Supabase dependency