RelayLab is a local-first experiment in autonomous software development.

Two Ollama-powered agents collaborate on the same isolated workspace:

• Builder — implements the requested app by proposing complete file writes/deletes.
• Reviewer — independently inspects the resulting workspace and either approves it or sends concrete feedback.
• Orchestrator — applies changes, performs deterministic static checks, logs every round, and stops at a hard iteration limit.

No cloud model API is required.

v0.1 experimental scaffold

The first milestone is deliberately constrained. The agents can create/edit files only inside one experiment workspace. They do not receive shell access in v0.1. This lets us observe the Builder → Reviewer feedback loop before adding more dangerous capabilities.

• Python 3.11+
• Ollama
• Local models:

ollama pull qwen2.5-coder:3b
ollama pull qwen3:4b

These lighter defaults are intended to keep the first experiment responsive on consumer hardware. Larger models can still be selected from the CLI.

Verify Ollama is running:

ollama list
curl http://localhost:11434/api/tags

git clone https://github.com/miflow13/RelayLab.git
cd RelayLab

python -m venv .venv
source .venv/bin/activate

python -m unittest discover -s tests -v

python relay.py \
  "Build a tiny single-page notes app with add, complete, and delete actions." \
  --name notes-v1

RelayLab will create:

experiments/notes-v1/workspace/   # app produced by the agents
runs/<run-id>/                    # full experiment transcript + metadata

task
  ↓
Builder
  ↓
structured file actions
  ↓
RelayLab applies changes
  ↓
static deterministic checks
  ↓
Reviewer
  ├─ changes_requested → Builder
  └─ approved → done

The loop stops after 6 rounds by default even if the Reviewer never approves.

The Builder must return JSON:

{
  "summary": "What changed",
  "files": [
    {
      "path": "index.html",
      "content": "<!doctype html>..."
    }
  ],
  "deletes": []
}

The Reviewer must return JSON:

{
  "status": "approved",
  "feedback": "Why this satisfies the task."
}

Ollama is requested to return JSON-formatted responses, and RelayLab validates the protocol before applying actions.

v0.1 intentionally does not give either model unrestricted computer access.

• Files are resolved against one experiment workspace.
• Absolute paths and .. path traversal are rejected.
• Model-authored code is not executed.
• Python source is syntax-compiled without running it.
• JSON files are parsed without running generated programs.
• Git operations and pushes are unavailable to the agents.
• Every agent response, applied action, check result, and review is logged.
• Runs have a hard maximum-round limit.

This is still experimental software rather than a hardened security boundary. Keep generated work disposable until stronger process/container isolation is added.

The previous 7B/8B defaults were intentionally reduced after early testing showed they could overwhelm a desktop when Ollama fell back heavily to CPU.

Override either from the CLI:

python relay.py "Build a pomodoro timer" \
  --name pomodoro \
  --builder-model qwen2.5-coder:7b \
  --reviewer-model qwen3:8b

RelayLab is meant to become an actual experiment, not just a chatbot demo. Each run records enough information to compare:

• rounds to approval
• reviewer rejection reasons
• deterministic check failures
• model combinations
• final artifacts
• whether a two-agent loop improves over a single-agent baseline

RelayLab/
├── relaylab/
│   ├── agents/
│   │   ├── builder.py
│   │   └── reviewer.py
│   ├── config.py
│   ├── jsonutil.py
│   ├── ollama_client.py
│   ├── orchestrator.py
│   └── workspace.py
├── prompts/
│   ├── builder.md
│   └── reviewer.md
├── experiments/
├── runs/
├── tests/
├── relay.py
└── pyproject.toml

• Local Ollama client
• Builder / Reviewer role separation
• Structured action protocol
• Workspace path isolation
• Hard iteration limit
• JSON run transcripts
• Deterministic static checks
• GitHub Actions unit tests
• Strong OS/container sandbox
• Controlled shell/tool calling
• Browser-based app verification
• Git checkpoints per accepted round
• Single-agent baseline mode
• Run comparison metrics
• Live web UI for watching both agents work

Experimental project. Add a license before distributing or incorporating third-party code.