Skillget

DQIII8

agent
SUMMARY

Works for you. Go outside and live. — AI orchestrator that auto-routes tasks to the cheapest model that solves them. 70% run free on local models. Self-auditing, self-improving, zero prompting skill needed. Built with vibe coding by a finance student. Your models, your data.

README.md

DQIII8

Tests License: MIT Python 3.10+ Platform Claude Code

Support on Ko-fi

DQIII8 makes any AI model smarter by injecting domain knowledge and restructuring prompts before they reach the model.

  • Routes prompts to the cheapest model that can handle the task (local → free cloud → paid)
  • Enriches prompts with relevant domain knowledge before the model sees them
  • Adapts prompt structure per model tier (small models get more scaffolding, large ones get raw data)
  • Autonomous operation via Telegram bot with 23 commands
  • 12 lifecycle hooks for Claude Code integration (auto-permissions, auto-commit, session tracking)
  • 17 slash skills for auditing, handover, security testing, and more

Quick Start

git clone https://github.com/senda-labs/DQIII8
cd DQIII8
bash install.sh

Requirements: Ubuntu 22.04/24.04 (or WSL2), Python 3.10+, 8 GB RAM. No GPU needed.

After installation:

# Configure your API keys
cp config/.env.example .env
nano .env

# Initialize the knowledge base
for d in applied_sciences formal_sciences natural_sciences social_sciences humanities_arts; do
    python3 bin/agents/knowledge_indexer.py --domain "$d"
done

# Start the bot
systemctl enable --now dqiii8-bot

# Verify
python3 -m pytest tests/test_smoke.py -q

How It Works

prompt → classify domain → retrieve knowledge → restructure prompt → route to model → response

The 8-step DQ pipeline classifies your prompt's domain, retrieves relevant knowledge chunks via hybrid search (vector + FTS5), restructures the prompt with tier-appropriate scaffolding, and routes to the cheapest model that can handle the task.

Tier Provider Cost When
C Ollama (local) $0 Code, debug, git
B Groq (free) $0 Research, analysis, writing
B+ GitHub Models $0 Fallback, long-context code
A Anthropic (paid) ~$0.01–0.05 Finance, architecture, orchestration
S Anthropic (paid) ~$0.15–0.50 System design, multi-agent coordination

The system always picks the cheapest tier that can handle the task.

Supported Models

  • Ollama (local, no API key): qwen2.5-coder:7b, bge-m3 (embeddings), any Ollama-compatible model
  • Groq (free): llama-3.3-70b-versatile — supports up to 9 API keys for round-robin rate limit management
  • GitHub Models (free, requires GITHUB_TOKEN): deepseek-v3, codestral-2501, gpt-4o-mini
  • OpenRouter: Any model via OPENROUTER_API_KEY
  • Anthropic: claude-sonnet-4-6, claude-opus-4-6 (via API key or Claude Code CLI OAuth)

Architecture

┌─────────────────────────────────────────────────────┐
│                    Entry Points                      │
│  Telegram /cc  │  CLI j cc  │  Director  │  Dashboard │
└───────┬─────────────┬───────────┬───────────┬────────┘
        │             │           │           │
        ▼             ▼           ▼           ▼
┌─────────────────────────────────────────────────────┐
│              DQ Pipeline (8 steps)                    │
│                                                       │
│  [1] Domain Classifier (keyword → embedding fallback) │
│  [2] Subdomain Classifier                             │
│  [3] Hierarchical Router (softmax, multi-level)       │
│  [4] Agent Selector (27 specialists, 5 domains)       │
│  [5] Knowledge Enricher (hybrid: vector+FTS5+graph)   │
│  [6] Confidence Gate (should enrich?)                 │
│  [7] Intent Amplifier (tier-specific prompt design)   │
│  [8] Stream Response (with fallback chain)            │
└─────────────────────────────────────────────────────┘

Claude Code Integration

DQIII8 deeply integrates with Claude Code through:

  • 12 lifecycle hooks — auto-permissions, session tracking, security scanning, auto-commit
  • 17 slash skills — /audit, /handover, /checkpoint, /mobilize, /red-team, and more
  • PermissionAnalyzer v5 — intelligent APPROVE/DENY/ESCALATE for every tool use
  • Auto-commit on session close — never lose work

Configuration

cp config/.env.example .env
nano .env
Variable Source Required
GROQ_API_KEY console.groq.com Yes (free)
GITHUB_TOKEN github.com/settings/tokens Recommended (free)
OPENROUTER_API_KEY openrouter.ai Optional
ANTHROPIC_API_KEY console.anthropic.com Optional
TELEGRAM_BOT_TOKEN @BotFather For Telegram UI
TELEGRAM_CHAT_ID Your Telegram user ID For Telegram UI

At minimum, add a GROQ_API_KEY (free) to enable Tier B. Tier C (Ollama local) works with no keys at all.

Telegram Bot Commands

View all 23 commands
Command Description
/cc <prompt> Execute via Claude Code (rate: 10/hr)
/status [project] System or project status
/loop [project] [cycles] Start autonomous loop
/stop Stop autonomous loop
/audit Full system health audit
/dq DQ pipeline metrics
/tasks List active tasks
/task <id> Task detail
/output <id> Task output
/kill <id> Terminate task
/score Model satisfaction scores
/logs Recent errors
/cc_status Claude Code status
/auth_status OAuth status
/auth_test Test authentication
/auth_update Refresh OAuth
/images [query] Generate images
/voice on|off Toggle TTS
/sandbox_run Run sandbox tester

Documentation

📖 Full System Guide — Complete reference with installation, commands, agents, hooks, pipeline details, and troubleshooting.

Updating

cd /root/dqiii8
bash update_dqiii8.sh

Or manually:

git pull origin main
pip install --break-system-packages -U crawl4ai pdfplumber docxtpl
for d in applied_sciences formal_sciences natural_sciences social_sciences humanities_arts; do
    python3 bin/agents/knowledge_indexer.py --domain "$d"
done
systemctl restart dqiii8-bot
python3 -m pytest tests/test_smoke.py -q

Contributing

See CONTRIBUTING.md.

Support the Project

If DQIII8 is useful to you, consider supporting development:

Support on Ko-fi

License

MIT — see LICENSE.

Reviews (0)

No results found