QonQrete v0.8.9-beta: The s00permode Release 🔥
“Don’t rebuild. Iterate. That’s the whole philosophy.”
What’s New in v0.8.9
We fixed the rebuild bug. But not with restrictions—with awareness.
The Universal File Rule
One rule. All cycles. No exceptions.
📁 File EXISTS in qodeyard?
→ MODIFY it
→ EXTEND it
→ Never recreate it
📄 File DOESN'T EXIST?
→ CREATE it
→ New modules welcome
That’s it. No “refinement modes”. No artificial restrictions. Just a simple rule that prevents the AI from starting over while keeping full creative freedom.
Cycles 2-7 can now:
- ✅ Create entirely new modules
- ✅ Add missing functionality
- ✅ Fix issues from reviews
- ✅ Extend existing code
But they won’t:
- ❌ Recreate main.py if it exists
- ❌ Rebuild the config system
- ❌ Start with “setup project root”
The Numbers
Cost Efficiency
| What | Before | After |
|---|---|---|
| 28 briqs reviewed | 28 API calls | 3 batches |
| Review cost | ~$1.40 | ~$0.01 |
| Full 7-cycle build | ~$50+ | ~$2.00 |
99% cost reduction on reviews. Not a typo.
Quality Improvements
| Metric | v0.8.0 | v0.8.9 |
|---|---|---|
| Syntax valid | 75% | 95% |
| Builds run | 60% | 85% |
| Functional output | 40% | 65% |
The Stack
# Your config for production builds
agents:
instruqtor:
provider: openai
model: gpt-4.1-nano # $0.10/1M - cheap planner
construqtor:
provider: gemini
model: gemini-2.5-flash # $0.30/1M - fast builder
inspeqtor:
provider: openai
model: gpt-4.1-mini # $0.40/1M - solid reviewer
options:
briq_sensitivity: 3 # 15-20 briqs per cycle
auto_cycle_limit: 7 # 7 iterations
mode: program
Total cost for a complex project: ~$2
How It Works
The Pipeline
┌─────────────────────────────────────────────────────────┐
│ CyQle N │
├─────────────────────────────────────────────────────────┤
│ │
│ 📋 InstruQtor │
│ └─→ Reads tasq.md (or previous reqap) │
│ └─→ Checks qodeyard (what exists?) │
│ └─→ Generates briqs (15-20 targeted tasks) │
│ │
│ 🔨 ConstruQtor (Interleaved) │
│ └─→ For each briq: │
│ └─→ Generate code │
│ └─→ [LoQal] Verify syntax │
│ └─→ Retry if failed (up to 3x) │
│ └─→ Write to qodeyard │
│ │
│ 🔎 InspeQtor (Batched) │
│ └─→ Group briqs into batches │
│ └─→ Review each batch (1 API call per batch) │
│ └─→ Meta-review consolidates │
│ └─→ Generate reqap for next cycle │
│ │
│ ⏸️ CheQpoint │
│ └─→ [Q]ontinue / [T]weaQ / [X]Quit │
│ │
└─────────────────────────────────────────────────────────┘
The Secret Sauce
LoQal Verification catches errors BEFORE the review:
[LoQal] Running validation...
[LoQal] ✅ Passed
Per-Briq Retry recovers from failures:
Attempt 1: Failed (API error)
Attempt 2: Success ✅
Batched Reviews save 90% on API calls:
Batch 1/3: 12 briqs → 1 API call
Batch 2/3: 12 briqs → 1 API call
Batch 3/3: 4 briqs → 1 API call
Get Started
# Clone it
git clone https://github.com/yourusername/qonqrete.git
cd qonqrete
# Configure your API keys
cp .env.example .env
# Edit .env with your keys
# Build the container
./qonqrete.sh init
# Run your first build
./qonqrete.sh run --auto
# Watch the magic happen
The Philosophy
QonQrete is built on three principles:
1. Local-First
Your code stays on your machine. Your API keys are yours. No cloud dependencies except the AI providers you choose.
2. File-Based Communication
Agents communicate through files, not memory. Everything is auditable. Every decision is logged.
3. Iterate, Don’t Rebuild
The Universal File Rule ensures every cycle builds on the last. No wasted work. No starting over.
What Can You Build?
QonQrete has successfully generated:
- Full-stack applications (FastAPI + React)
- CLI tools (Python, Go, Rust)
- Infrastructure code (Docker, Terraform)
- Complex systems (C2 frameworks, multi-service architectures)
- Test suites (pytest, unit tests, integration tests)
If you can describe it in a tasq.md, QonQrete can build it.
Community
- GitHub: github.com/qonqrete
- Discord: Coming soon
- Twitter: @qonqrete
License
MIT. Build whatever you want. Just don’t be evil.
QonQrete v0.8.9-beta
The local-first, multi-agent AI construction loop.