v1.2.0

QonQrete

The First 100% File-Based

Local-First Secure AI Dev Construction Yard

GET QONQRETE FOR YOUR IDE

🏗️ WHAT IS QONQRETE?

Imagine having a tiny software development team living inside a secure box on your computer. You give them a task, they plan it → build it → review it, and keep improving until it's done. All while you keep full control, your code never leaves your machine, and you can see exactly what they're doing.

THE PIPELINE

📝

tasQ.md YOUR INPUT

You write what you want in plain English. "Build a REST API with user authentication" - that's it!

🏠 Like: Telling an architect "I want a 3-bedroom house with a pool"

🎯

TasqLeveler ENHANCER 🧠 AI

Automatically supercharges your task with dependency graphs, test criteria, and success metrics.

🏠 Like: An architect adding "also needs plumbing, electrical, and building permits"

📋

InstruQtor THE PLANNER 🧠 AI

Breaks your task into small, ordered steps called "briQs". Each briQ is one specific thing to build.

🏠 Like: Project manager creating work orders: "First foundation, then walls, then roof..."

🔨

ConstruQtor THE BUILDER 🧠 AI

Goes through each briQ and writes actual code. Builds file by file, following the plan exactly.

🏠 Like: The construction workers actually building according to the blueprints

🔍

InspeQtor THE REVIEWER 🧠 AI

Reviews all the generated code like a senior developer. Finds bugs, suggests improvements for the next cycle.

🏠 Like: Building inspector checking everything meets code and works properly

✋

CheQpoint HUMAN CONTROL ⚡ FREE

The system pauses. You review what was built. Continue, tweak, or stop - you're always in control.

🏠 Like: Client walkthrough - "I love it! But can we make the kitchen bigger?"

🔄

Next CyQle ITERATE

The review becomes the new task. The loop continues, each cycle improving the code until it's perfect.

🏠 Like: Renovation cycles until the house is exactly what you dreamed of

THE SECRET SAUCE: LOCAL AGENTS

These run on YOUR machine with ZERO API costs. They make the AI agents way more efficient:

🦴

QOMPRESSOR

Strips code down to just signatures and structure ("skeletons"). The AI sees the shape without reading every line.

→ 96% less tokens sent to AI

🔍

QONTEXTOR

Maps all functions, classes, and how they connect. Creates a "brain map" of your codebase.

→ AI knows what exists before asking

🌀

QONTRABENDER

Smart caching. Only sends what changed. Remembers what the AI already knows.

→ Near-zero repeated work

✅

LOQAL VERIFIER

Checks syntax and imports BEFORE sending to review. Catches obvious errors for free.

→ No paying AI to find typos

WHY THIS MATTERS

🔒

SECURE BY DESIGN

Everything runs in a Docker container (the "Qage"). AI-generated code can't touch your system.

💰

25X CHEAPER

Local preprocessing means you send way less to paid APIs. A $25 project becomes $1.

👀

FULLY TRANSPARENT

Every step writes to markdown files. You can read exactly what each agent thought and did.

🏠

LOCAL FIRST

Your code stays on YOUR machine. No cloud dependency. Works offline (except for AI calls).

⚡ SEE IT IN ACTION

$ ./qonqrete.sh run --auto

Get started View on GitHub

PREDICTABLE. TIGHT. ANTI-DRIFT.

Vibes don't ship. Contracts do. QonQrete v1.0.4 enforces rules that keep your builds convergent.

🛡️

QONTRACTGUARD

Validates QONTRACT invariants and rule sets before and after critical steps. If a constraint breaks, the next cyQle gets corrective guidance automatically.

→ Contracts > Vibes

🔄

INSPEQTOR + GUARD LOOP

InspeQtor reviews the code. QontractGuard validates invariants. Together they form a convergent loop that reduces drift across cyQles.

→ Each cyQle gets tighter, not looser

📐

DETERMINISTIC RUNS

Same input + same contracts = same enforcement. No stochastic surprises from model mood swings. Predictability is a core feature, not a side effect.

→ Stability as a guarantee

USERGATE: PER-CYQLE CONTROL

Full-auto is great until it isn't. UserGate mode gives you a steering wheel at every CheQpoint.

✋

CHEQPOINT GATING

After every cyQle, the system pauses. You review the ReQap, inspect artifacts, and decide: continue, tweak direction, or stop.

→ Human-in-the-loop, every loop

✏️

EDIT THE REQAP

Don't just approve or reject — edit the review summary before the next cyQle starts. Steer the build with surgical precision.

→ Adjust direction mid-flight

⚡

AUTO vs. GATED

Run fully autonomous with --auto or user-gated with --user. Switch between modes per run. No config rewrite needed.

→ Control when you want it, speed when you don't

WHY FILE-BASED (AUDIT TRAIL + PRIVACY)

Every decision, every artifact, every reasoning step — on disk, inspectable, yours. No black boxes.

📂

FULL ARTIFACT TRAIL

Every run produces: bloq.d/ (skeletons), qontext.d/ (semantic maps), reqap.d/ (reviews), struqture/ (decision logs), and meta.yaml (run metadata).

→ See exactly why decisions were made

🔐

KEEP CODE OUT OF CLOUD AI

Feed agents only qontext.d/, bloq.d/, tasq*, and qontract* artifacts instead of your full repo. Your source stays local.

→ Full intelligence, zero code exposure

🔁

REPRODUCIBLE & DEBUGGABLE

With everything on disk, you can replay, diff, and debug any run. No "worked yesterday, broke today" mysteries.

→ Same files, same outcome

Why QonQrete?

Your context lives on disk, not in a black-box mood swing

No silent memory wipes, no "sorry I forgot." QonQrete keeps reasoning, structure, and history locally, inspectable, reproducible — unlike OpenAI, Google (Gemini), or Anthropic.

No 90k-token hallucination cliff

QonQrete doesn't gamble on giant context windows like Alibaba Cloud (Qwen). It prepares context deterministically before the model ever speaks. If it's not on disk, it doesn't pretend it knows it.

LLMs are engines — QonQrete is the brain

Models come and go. Memory, reasoning, and decisions stay under your control, not trapped in provider-side chat state like with DeepSeek or the rest of the big cloud crew.

Reproducible runs, not vibes-based answers

Same input, same files, same outcome. No "worked yesterday, broke today" because the model woke up on the wrong side of the datacenter.

Maximum locality, minimum dependency

Pull the brains as close to the metal as possible: local files, local agents, optional local models. Cloud LLMs become replaceable text generators, not single points of cognitive failure.