Disha

Changelog

All notable changes to the Disha AGI Platform are documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

GNN overfitting resolved — Test accuracy improved from 7.2% to 75% on synthetic graph. Root causes: random labels, sequential split, insufficient regularization.
- ai-platform/backend/graph_ai/train.py: Labels now derived from node features instead of random assignment; train/test split uses shuffled permutation; added early stopping with patience-based checkpoint restoration.
- ai-platform/backend/graph_ai/models.py: Added BatchNorm1d to GCN encoder; increased dropout from 0.3 to 0.5; increased weight decay to 5e-4.
graph_ai import dependency — graph_ai/__init__.py now uses lazy __getattr__ for GraphExporter to avoid requiring pydantic_settings at import time. Models and trainer can be imported directly without the full dependency chain.
continuous_train.py — Updated GNN training section to use improved model architecture and synthetic graph generation with feature-derived labels.

GNN node classification: 150 epochs, 98.1% train / 75.0% test accuracy (synthetic), ~99.8%/99.8% (knowledge graph)
GNN link prediction: 200 epochs, loss 1.299
RL PPO: 400 episodes, reward 22.24 (±3.23)

LEARNING_LOG.md — Bumped to v3.2.0; updated version history, achievements, training metrics, demerits (GNN overfitting marked resolved), audit checklist
README.md — Updated version badge to v3.2.0; corrected stats (3,700+ files, 452K+ LoC, 19 Dockerfiles, 9 CI workflows, 13 test files); added v3.2.0 section
WIKI.md — Added Section 21 documenting v3.2.0 GNN fixes; updated GNN section 4.8 with new architecture
CHANGELOG.md — Added v3.2.0 entry

Orchestrator DNS relationship logic — Fixed _build_relationships() in ai-platform/backend/app/agents/orchestrator.py to validate target entity type for DNS records. Previously, DNS records created spurious RESOLVES_TO edges to any entity type; now restricted to host and domain targets only.
Quality score overflow — Fixed learning_controller.py credibility scoring that could exceed the documented 0–25 range (max was 30 due to uncapped cred_bonus). Now properly capped with min(25, ...).

server.json — Replaced incorrect monster-codemaster name and Monster/claude-code repository URL with correct disha-mcp and Tashima-Tarsh/Disha.
mcp-server/server.json — Same corrections as above.
mcp-server/package.json — Fixed package name, mcpName, repository URL, homepage, bugs URL, author, and bin entry to reflect the Disha project.

CONTRIBUTING.md — Fixed clone URL from Monster/claude-code to Tashima-Tarsh/Disha, added Python prerequisites and module setup instructions.
USAGE_GUIDE.md — Complete rewrite to accurately document Disha’s CLI, AI platform, decision engine, historical strategy, cyber defense, MCP server, training commands, continuous learning, sentinel monitoring, and troubleshooting.
README.md — Updated to v3.1.0, added comprehensive review section documenting all bug fixes and improvements.
CHANGELOG.md — Replaced stub content with full version history from v1.0.0 to v3.1.0.
WIKI.md — Added v3.1.0 review and bug fix documentation.

Full codebase audit: 2,477 source files, 9 CI workflows, 22 test files, 13 Dockerfiles
Verified no merge conflict remnants across entire repository
Verified all Python module imports resolve correctly with no circular dependencies
Verified all 7 AI agents properly inherit from BaseAgent and connect through orchestrator
Verified RL environment STATE_DIM=12 matches policy network input dimensions
Verified GNN models gracefully fall back when torch_geometric is unavailable

Universal knowledge bases — 8 domains: Physics, Mathematics, Computing, Chemistry (all 118 elements), Law & Politics, Cybersecurity, Innovation & Future Tech, Historical Strategy
Cross-domain knowledge engine (scripts/knowledge_engine.py) — loads all 8 domains and builds cross-domain knowledge graphs
Continuous training pipeline (scripts/continuous_train.py) — RL, GNN, decision engine training with online/offline data
Sentinel monitoring system (scripts/sentinel/) — threat_monitor.py (5 feeds), model_orchestrator.py (5 AI providers), guardian.py (auto-heal + health checks)
Auto-learning bot (auto_learning/) — RAG pipeline, LLaMA fine-tuning, multi-agent coordination, advanced reasoning
Disha Mythos workflow — scheduled learning agent execution

Quantum Physics Intelligence (quantum-physics/) — Qiskit circuit simulator, space APIs, physics engines
Decision Framework (decision-framework/) — 4 agents (Political, Legal, Ideology, Security) with FAISS retrieval
Decision Engine (decision-engine/) — Refined multi-agent reasoning with consensus voting
100% open-source API migration — Replaced all paid/stub APIs with free alternatives (ip-api, HackerTarget, Feodo Tracker, OpenStreetMap, local Whisper, LLaVA via Ollama)
Multimodal AGI — Vision agent (GPT-4o/LLaVA), Audio agent (Whisper), cross-modal fusion

Core CLI Engine — TypeScript + Bun + React/Ink with 40+ tools and 50+ commands
AI Intelligence Platform — FastAPI backend with 7 specialized agents (OSINT, Crypto, Detection, Graph, Reasoning, Vision, Audio)
Reinforcement Learning Engine — PPO policy with 12-dim state, 8 actions, prioritized experience replay
Graph Neural Networks — GCN encoder, link predictor, graph classifier
Cyber Defense System — Cowrie/Dionaea/OpenCanary honeypots, PyTorch threat classifier
Historical Strategy Intelligence — 32+ conflicts, Random Forest + MLP classifier, scenario simulation
Web Dashboard — Next.js 14 + Tailwind CSS with 78+ components
MCP Server — Model Context Protocol v1.29 server with STDIO/HTTP/SSE transports
Docker infrastructure — 13 Dockerfiles, multi-stage builds
CI/CD pipelines — 9 GitHub Actions workflows (CI, AI Platform, Decision Engine, Cyber Defense, CodeQL, Sentinel, Continuous Training, Modules, Mythos)

This site is open source. Improve this page.