Coding Tentacle 11

🐙 Secure Self-Learning Repair Agent — Droste Fusion
The only agent with Safety VETO + Causal Code Graph + Self-Learning

---

Why Coding Tentacle?

The Problem: AI coding agents — OpenCode, Claude Code, Codex, Copilot — generate code with zero safety guarantees. They can emit eval(user_input), DROP TABLE, or rm -rf / without hesitation. They use grep, not code graphs. They never learn from mistakes.

The Solution: Coding Tentacle is a Cyber-Safety Layer that sits between bug reports and fix engines. It doesn't generate code — it controls what code gets generated.

Metric	CT + OpenCode + Droste	Baseline (grep-only)
Repair Accuracy	9/10 (90%)	3/10 (30%)
Security Blocks	2 (eval, API key)	0
Causal Context	4,697 edges	0
Learning Retention	92 experiences, 23 REFLECTION entries	None

3× improvement over grep-only baselines. Security VETO catches dangerous patterns before execution.

Architecture

                          ┌──────────────────────────────────┐
  Bug Report ────────────►│        CODING TENTACLE            │
                          │                                   │
                          │  ┌─ Classification (18 types) ─┐  │
                          │  │  SecurityBrain VETO         │  │
                          │  │  Droste Causal Graph        │  │
                          │  │  ReflectionEngine           │  │
                          │  │  4-Engine Router            │  │
                          │  └─────────────────────────────┘  │
                          │                                   │
                          │   APPROVE / BLOCK / REQUEST_MORE  │
                          └──────────────┬───────────────────┘
                                         │ controls
  ┌──────────────────────────────────────▼────────────────────┐
  │                    FIX ENGINES                             │
  │                                                           │
  │   OpenCode (deepseek-v4)  │  Claude Code (2.1.86)         │
  │   Ollama (granite3.2)     │  Codex (disabled)             │
  │                                                           │
  │   Sandbox → Safety Scan → Impact Analysis → Approval       │
  │                                                           │
  │   CANNOT act without CT approval. CANNOT bypass safety.    │
  │   CANNOT access files directly. CANNOT auto-commit.        │
  └───────────────────────────────────────────────────────────┘

CT + Droste Fusion

CT uses Droste — a causal code graph engine — to give every fix engine structural context. Instead of blind grep, the LLM sees which functions are causally connected to the bug.

CODING TENTACLE                         DROSTE v1.1.6
━━━━━━━━━━━━━━━━━━              ┏━━━━━━━━━━━━━━━━━━━━━━━┓
Classifier (18 types)           ┃ 4,187 symbols          ┃
SecurityBrain (VETO)            ┃ 4,697 causal edges     ┃
EngineRouter (4 engines)        ┃ 98% budget efficiency  ┃
ReflectionEngine                ┃ "Which functions are   ┃
BLM + EngineLearning            ┃  causally connected?"  ┃
━━━━━━━━━━━━━━━━━━              ┗━━━━━━━━━━━━━━━━━━━━━━━┛
         │                              │
         └────────── FUSION ────────────┘
                    │
          Causal Code Context in Engine Prompt

Full Pipeline

BUG → Classifier → SecurityBrain(VETO) → REFLECTION RETRIEVAL
    → Droste Context → EngineRouter → OpenCode / Claude / Ollama
    → Safety Scan → ImpactAnalyzer → SkepticBrain
    → Sandbox → Approval → BLM → REFLECTION ENGINE → Consolidator

Comparison

Feature	CT	Claude Code	Codex	Copilot	Devin
Safety VETO (5-layer)	✅	❌	❌	❌	❌
Causal Code Graph	✅	❌ grep	❌ grep	❌ grep	❌ grep
Self-Learning (BLM)	✅	❌	❌	❌	❌
Open Source	✅	❌	❌	❌	❌
Local / No Cloud	✅	❌	❌	❌	❌
Multi-Engine Routing	✅	❌	❌	❌	❌
Immutable Evidence Trail	✅	❌	❌	❌	❌
No Auto-Apply	✅	❌	❌	⚠️	⚠️

CT is the only system with all 8 features.

Engines

Engine	Status	Type
OpenCode (deepseek-v4)	Primary	Local, free
Claude Code (2.1.86)	Active	Top-tier
Ollama (granite3.2)	Fallback	Offline
Codex (GPT-5.x)	Disabled	API required

Learning System V2

Run 1: Bug → Fix → REFLECTION: "Why did this work?"
Run 2: Bug → BLM: "I've seen this!" → LEARNED LESSONS enrich prompt
Run N: 92 experiences, 23 REFLECTION entries, trust calibrated

Quick Start

# Clone and install
git clone https://github.com/nessos666/coding-tentacle.git
cd coding-tentacle
pip install -e .

# Verify installation
ct version

# Repair a bug
ct repair --title "NullPointer in payment.py" --body "NoneType has no attribute 'process'"

Requirements: Python 3.10+ • Droste v1.1.6 • OpenCode CLI • No API keys • No cloud

Project Status

Metric	Value
Version	11.0.0
Modules	27 production
Lines of Code	8,405
Test Suite	14/14 passing
Regression	10/10 passing
Checkpoints	28
Archived Modules	34
CI/CD

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines and CODE_OF_CONDUCT.md for our community standards.

License

MIT — Built by David Miko + Hermes Agent. © 2026.