Autonomous Agentic Coding Bot - A local agent that is allowed to be silent and can’t hurt your machine.

This repo contains a multi-agent pipeline (Orchestrator → Planner → Worker → Reviewer) with project-specific specs stored under projects/<project>/project.yaml. Runs, logs, and memory stay inside the project tree.

What LLM-Sandbox is NOT

• An AI model
• A finished product
• A way to be lazy

Prerequisites

• Docker & Docker Compose
• .env with:
  • OPENAI_API_KEY (required)
  • CLOUDFLARE_API_TOKEN (if doing Cloudflare work)
  • GIT_USER_NAME and GIT_USER_EMAIL (optional; used to set per-project git author)

Build and start the container

docker compose up -d --build

Run a project

1. Kick off a run (the orchestrator will scaffold the project if missing, including project.yaml):
2. From inside the container (or via docker exec):

python -m agents.orchestrator -n <project_name> "<goal>"

Examples:

docker exec -it llm-sandbox bash
python -m agents.orchestrator -n franchisetalk "Publish a hello world website to <my domain>.com"

Project output

• projects/<project>/output/ – generated deliverables (apps, sites, assets, campaigns)
• projects/<project>/runs/ – run summaries
• projects/<project>/logs/ – agent logs
• projects/<project>/state.json – runtime session state (auto-created)
• memory/ – memory index (across projects)
• projects/<project>/project.yaml – required project spec

Project scaffolding, git, and state

• python -m agents.orchestrator uses initialize_project to create projects/<name>/, output/, project.yaml, and state.json.
• project.yaml is auto-generated if missing using the richer schema: project_id, name, goal, description, repo{url,default_branch,ssh_remote_name}, default_execution_mode, rdm_agents{planner_id,worker_id,qa_id,analyst_id}, steps, metadata{created_at,tags}.
• Each project gets its own git repo on first run:
  • Initialized on branch main, with local git config from GIT_USER_NAME / GIT_USER_EMAIL if set.
  • Project-level .gitignore ignores state.json, logs/, runs/ (output is left trackable).
  • No remotes are added or pushed; you can add one later inside projects/<name>.

Notes

• The worker is sandboxed to projects/<project> via WORKSPACE_ROOT; generated files belong under projects/<project>/output/ (create if missing).
• Use python -m agents.<agent> to test individual agents if needed.
• To inspect a new project repo: cd projects/<project> && git status (initial commit contains project.yaml, .gitignore, and an output/.gitkeep placeholder).