agentic-coding-assistant

An agentic code generation and review pipeline built with Quarkus and LangChain4j. A plain-text spec goes in — reviewed, security-checked Java code comes out.

Local LLM. Five agents. Two self-correcting loops. Your code never leaves your machine.

---

How It Works

The pipeline chains five specialized agents in a fixed sequence:

Agent	Role	Scope key
Coder	Writes initial code from the spec	writes code
Reviewer	Checks quality checklist — APPROVED or FAILED: <reason>	writes review_status
Editor	Rewrites code on failure, returns to Reviewer	writes code
SecurityReviewer	OWASP-style security pass	writes security_status
SecurityEditor	Patches security issues, returns to SecurityReviewer	writes code

Review loop (Reviewer ↔ Editor) runs up to 3 iterations until APPROVED.
Security loop (SecurityEditor ↔ SecurityReviewer) runs until APPROVED.

All agents communicate exclusively through a shared agentic scope — a key/value store. No agent holds a reference to any other agent.

---

Prerequisites

Requirement	Version
Java	25
Maven	via ./mvnw wrapper
Ollama	latest — ollama.ai
Model	qwen2.5-coder:3b (default)

Pull the model before starting:

ollama pull qwen2.5-coder:3b

---

Quick Start

# Start with Ollama (default)
./mvnw quarkus:dev

# Start with an OpenAI-compatible endpoint (e.g. LM Studio)
./mvnw quarkus:dev -Popenai

Or use Docker Compose — see Docker below.

Open http://localhost:8090, enter a spec such as:

Write a Java method that validates an email address

The status log shows each agent as it runs:

› Coder is working...
› Coder finished: public static boolean isValidEmail(String email) {...
› Reviewer is working...
› Reviewer finished: FAILED: No Javadoc, missing null check on input...
› Editor is working...
› Editor finished: /** Validates whether the given string is a valid emai...
› Reviewer is working...
› Reviewer finished: APPROVED
› SecurityReviewer is working...
› SecurityReviewer finished: APPROVED

---

Configuration

LLM provider is selected via application.properties or Maven profile. No code changes required to switch.

Ollama (default)

quarkus.langchain4j.chat-model.provider=ollama
quarkus.langchain4j.ollama.base-url=http://localhost:11434
quarkus.langchain4j.ollama.chat-model.model-id=qwen2.5-coder:3b
quarkus.langchain4j.ollama.timeout=PT300S

OpenAI-compatible endpoint (-Popenai)

%openai.quarkus.langchain4j.openai.base-url=http://localhost:1234/v1
%openai.quarkus.langchain4j.openai.api-key=lm-studio
%openai.quarkus.langchain4j.openai.chat-model.model-id=local-model

Override via environment variables — see .env.template.

---

Endpoints

URL	Description
http://localhost:8090/	Web UI
ws://localhost:8090/coding	WebSocket — send spec as plain text, receive JSON messages
http://localhost:8090/topology	Topology Report — HTML execution trace of the last pipeline run

WebSocket message format

{ "type": "connected", "content": "Pipeline ready." }
{ "type": "status",    "content": "Coder is working..." }
{ "type": "result",    "content": "<generated code>" }
{ "type": "error",     "content": "<error message>" }

---

Topology Report

After each pipeline run, GET /topology returns an HTML report of the full execution trace — which agents ran, how many loop iterations occurred, and how long each step took.

Useful for debugging slow pipelines and for auditing what the agents actually did — relevant when AI-generated code goes into a team review workflow.

---

Project Structure

src/main/java/de/markserver/agentic/coding/
  ├── boundary/
  │     ├── CodingWebSocket.java       @WebSocket(path="/coding")
  │     └── TopologyResource.java      GET /topology
  └── control/
        ├── CodingAgents.java          5 agent interfaces
        ├── CodingPipeline.java        pipeline builder (@ApplicationScoped)
        └── WebSocketAgentListener.java AgentListener → WebSocket streaming

src/main/resources/
  ├── application.properties
  └── prompts/                         externalized prompt templates

---

Docker

A pre-built image is published to the GitHub Container Registry on every push to main — built with Quarkus Jib, no Docker daemon required in CI.

docker pull ghcr.io/mgoericke/agentic-coding-assistant:latest

Docker Compose — Ollama (local LLM, no API key)

Pulls qwen2.5-coder:3b automatically on first run (~2 GB).

docker compose up

Docker Compose — OpenAI-compatible endpoint

Works with LM Studio, the OpenAI API, or any compatible endpoint.

# LM Studio running on your host machine (default)
docker compose -f docker-compose.openai.yml up

# OpenAI API
OPENAI_BASE_URL=https://api.openai.com/v1 \
OPENAI_API_KEY=sk-... \
OPENAI_MODEL=gpt-4o \
docker compose -f docker-compose.openai.yml up

Build the image locally

./mvnw package -Dquarkus.container-image.build=true

---

Key Dependencies

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-agentic</artifactId>
    <version>1.13.1-beta23</version>
</dependency>
<dependency>
    <groupId>io.quarkiverse.langchain4j</groupId>
    <artifactId>quarkus-langchain4j-ollama</artifactId>
</dependency>
<dependency>
    <groupId>io.quarkus</groupId>
    <artifactId>quarkus-websockets-next</artifactId>
</dependency>

langchain4j-agentic provides the Sequence, Loop, Parallel, and Conditional pipeline primitives. Agents and pipelines are built programmatically per request — no hidden CDI wiring.

---

Running Tests

# Unit tests
./mvnw test

# Integration tests
./mvnw verify -DskipITs=false

---

Roadmap

• Test coverage — @QuarkusTest smoke tests for WebSocket and /topology endpoint
• Feedback loop — send corrections back into a second pipeline run
• TestWriter agent — generate unit tests as a sixth pipeline step
• Picocli CLI boundary — run the same pipeline from terminal or CI
• Human in the Loop — pause after first Reviewer pass for manual approval
• Domain-specific skills — swap agent prompts for document review, PR summaries, or requirements analysis

---

Tech Stack

Runtime	Quarkus 3.35.1
Language	Java 25
Agentic pipeline	langchain4j-agentic 1.13.1-beta23
LLM providers	Ollama · OpenAI-compatible (LM Studio)
WebSocket	quarkus-websockets-next
REST	quarkus-rest-jackson