⚡ ANF — Autonomous Native Forge

Drop a PRD. Get working software. See what happens in between.

Autonomous Native Forge (ANF) is a 5-agent software factory that reads technical documents (PRD, Sprint, Spec) and autonomously produces working, production-ready software. It runs entirely locally and has zero npm dependencies in its core architecture.

• No cloud. No vendor lock-in. No mandatory API keys.
• Pure Node.js. Only node:http, node:fs, node:path, node:events.
• Every LLM error, every retry, and every steering decision is recorded in DEVLOG.md and llm_communication.log.
• 24/7 Autonomous Telemetry tracking performance, cost, and progress in anf_system_report.md.

---

🚀 Quick Start

# 1. Verify NIM/LLM connection
npm run test-nim

# 2. Start the factory (spawns all agents + telemetry)
npm run forge

# 3. Open the dashboard in another terminal
npm run dashboard
# → http://localhost:3000 (auto-refreshes every 5 seconds)

# 4. Drop your project's PRD → Architect will discover it automatically
mkdir -p docs/reference/YOUR_PROJECT_NAME
# Place your prd.md file there

---

✨ Enterprise & Sovereign Capabilities (V4.6)

ANF is not just a coding tool; it is a Sovereign Industrial System designed for high-availability autonomous engineering.

🛡️ Sovereign Architecture (Self-Healing)

The system is equipped with a Self-Correction Protocol that allows it to repair its own brain (core agent logic) without human intervention.

• Sentinel Watchdog: Actively monitors agent heartbeats and process health.
• Quarantine Logic: Detects deterministic "Crash Loops" (3 crashes in 5 mins) and isolates the failing component.
• Autonomous Self-Patching: The Architect agent analyzes the failing agent's crash logs, identifies logic errors (ReferenceErrors, TypeErrors), generates a corrective patch, and overwrites the agent's source code to restore the factory line.

📜 Immortal Logging & Auditability

Data is the lifeblood of AI improvement. ANF implements a Zero-Deletion Policy.

• Permanent Audit Trail: Every task ever processed (Success or Failure) is kept permanently in queue/done/ and queue/error/.
• Timestamped Log Rotation: Massive communication logs are never overwritten. When llm_communication.log exceeds 100MB, it is automatically archived with an ISO timestamp, creating a forever-traceable history of every token generated by the forge.

🚀 Massive Context & Intelligence Scaling

Optimized for the NVIDIA Blackwell (GB10) architecture, ANF pushes the boundaries of model awareness.

• Project-Wide Awareness: Scaled tree visibility to 500+ files.
• Deep File Inspection: Increased file-read capacity to 10,000 characters per file, preventing "hallucinated imports" caused by truncated code.
• Context Sentinel: Watchdog detects "Context Overflow" (4096+ tokens) in real-time and resets agents to maintain reasoning precision.

📊 Industrial Telemetry & Quality

• 24/7 Telemetry Daemon: Real-time tracking of LoC/min, GPU thermal/power efficiency, and MTBF (Mean Time Between Failures).
• 5-Layer QA Gate: Every line of code passes through:
  1. Syntax Validation (Native Node.js)
  2. Static Security Scanning (Security Guardrails)
  3. Docker Sandbox Execution (Isolated Alpine container)
  4. PRD Compliance Check (Architect Review)
  5. Peer Review Consensus (Cost vs Performance analysis)

🌿 Advanced Workflow

• Sprint Branch Workflow: Automated pushing to feature/sprint-sN and opening Pull Requests to main upon completion.
• High-Concurrency Coder: Maximizes GPU batching by executing multiple NIM API calls simultaneously (up to vault.concurrency.CODER).
• Webhooks & Notifications: Integrated event-driven webhooks for TASK_FAILED, SPRINT_COMPLETE, and PR_OPENED.

---

🧠 LLM Architecture & Configuration

ANF natively integrates with the NVIDIA NIM OpenAPI format (/v1/chat/completions), specifically optimized for handling massive reasoning models like Nemotron-3-Super-120B.

Configuration — config/vault.json

{
  "global": {
    "nim_host": "localhost",
    "nim_port": 8000,
    "nim_protocol": "http",
    "nim_enable_thinking": true,
    "nim_reasoning_budgets": {
      "ARCHITECT": 16384,
      "CODER": 4096,
      "TESTER": 256
    },
    "concurrency": {
      "ARCHITECT": 1,
      "CODER": 3,
      "TESTER": 2
    }
  }
}

---

📁 Project Structure

AutonomousNativeForge/
├── agents/                    # Core AI Logic (Sovereign Core)
│   ├── watchdog.js            # Sentinel & Self-Healing Trigger
│   ├── architect.js           # Orchestrator & Logic Repairman
│   ├── coder.js               # Concurrent Code Generator
│   ├── tester.js              # 5-layer QA Gate
│   └── telemetry.js           # 24/7 System Health Daemon
├── queue/                     # Atomic IPC Messaging System
│   ├── inbox/                 # Incoming tasks (Atomic Writes)
│   ├── processing/            # Locked active tasks
│   └── done/                  # Immortal Task History
├── src/                       # Generated Source Code
├── DEVLOG.md                  # Human-readable completion logs
└── llm_communication.log      # Raw token audit trail (Auto-rotated)

---

👨‍💻 Author

Turgay Savacı — Software Developer, 15+ years in IT, specializing in Software Engineering & Autonomous Systems.

Cloud is convenient. Local is free. Sovereign is immortal.