🏗️ BanglaDev

24/7 autonomous development agent orchestrator — repository-scoped, local-LLM-first.

BanglaDev is a stripped-down, dev-focused fork of Paperclip AI. It drops the "zero-human company" metaphor and replaces it with a pre-built development team that lives inside your repository, powered entirely by local LLMs.

You point it at a repo. It understands the codebase. It picks up tickets. It writes code, runs tests, opens PRs. You review and merge.

---

Why BanglaDev?

Paperclip is brilliant orchestration — but it's designed for multi-company, multi-domain, frontier-model operations. If you want a strictly dev-focused agent team running on your own GPU hardware against your existing repos, you need something leaner.

What was removed:

• 7 frontier adapters (Claude, Codex, Cursor, Gemini, OpenCode, OpenClaw, Pi)
• Multi-company isolation, org chart visualization, branding system
• Company portability (import/export), feedback system, finance tracking
• ~1MB of TypeScript removed

What was added:

• aider-local adapter — wraps aider for code editing via local LLMs
• vllm-direct adapter — raw OpenAI-compatible chat for planning/review agents
• Repo context engine — auto-detects language, framework, test runner, file structure
• Pre-built 6-agent dev team with role-specific system prompts
• Dev-specific skills: git-ops, test-runner, code-review, codebase-index, deploy
• bangladev init — one command to bootstrap a dev team in any repo

---

The Dev Team

Agent	Role	Adapter	Heartbeat
👑 Lead Dev	Orchestrator — decomposes goals, delegates, reviews	vllm_direct	5 min
🏗️ Architect	System design, dependency analysis, refactoring	vllm_direct	10 min
⚙️ Backend Engineer	APIs, database, business logic, server code	aider_local	2 min
🎨 Frontend Engineer	UI components, styling, client-side logic	aider_local	2 min
🧪 QA Engineer	Test writing, coverage analysis, regression tests	aider_local	3 min
🚀 DevOps Engineer	CI/CD, deployment, build optimization	aider_local	10 min

The Lead Dev and Architect think (vllm_direct — direct LLM chat with tool calling). The engineers code (aider_local — aider CLI with git integration).

---

Supported Local Models

Model	Size	Use Case
MiniMax-M2.5 (NVFP4)	230B	Primary agentic model (TP=2 on dual DGX Spark)
Qwen3-30B-A3B	30B MoE	Fast orchestrator/router
Super Nemotron Nano	~8B	Lightweight task execution
Qwen2.5-Coder-32B	32B	Strong single-GPU coder
DeepSeek-Coder-V3	236B	Alternative heavy coder
Devstral Small	~24B	Specialized coding model

All accessed via OpenAI-compatible /v1/chat/completions — works with vLLM, Ollama, llama.cpp server, or any compatible endpoint.

---

Quick Start

# 1. Install
npm install -g bangladev

# 2. Go to your repo
cd /path/to/your/repo

# 3. Initialize (auto-detects repo, prompts for LLM endpoint)
bangladev init --endpoint http://localhost:8000/v1 --model minimax-m2.5

# 4. Start your LLM server (if not already running)
vllm serve minimax-m2.5 --port 8000 --gpu-memory-utilization 0.75

# 5. Start the dev team
bangladev start

# 6. Create a task
bangladev task create "Add user authentication with JWT"

# 7. Watch the dashboard
bangladev dashboard
# → http://localhost:3100

---

How It Works

Heartbeat Loop

Every agent runs on a heartbeat schedule. Each heartbeat:

1. Agent wakes up
2. Checks its task queue
3. Checks out the highest-priority task (atomic — no double-work)
4. Reads the codebase context injected by the repo context engine
5. Does work (plan, code, test, review — depending on role)
6. Commits changes, updates task status
7. Exits

Repo Context Engine

On every heartbeat, agents receive auto-detected context about the repository:

• Language and framework detection
• File tree (top 3 levels)
• Git state (branch, dirty files, recent commits)
• Test runner and build commands
• Entry points and dependency graph

This means agents understand the codebase before writing a line of code.

Task Flow

User creates goal: "Add user authentication"
    ↓
Lead Dev decomposes into tasks:
    ├── BD-1: Design auth architecture (→ Architect)
    ├── BD-2: Implement JWT middleware (→ Backend)
    ├── BD-3: Create login/signup pages (→ Frontend)
    ├── BD-4: Write auth integration tests (→ QA)
    └── BD-5: Add auth to CI pipeline (→ DevOps)
    ↓
Each agent picks up their task, works on a feature branch, commits
    ↓
Lead Dev reviews completed tasks, approves or requests changes
    ↓
Tasks merge to main

Git Discipline

• Feature branches per task: bangladev/{role}/BD-{number}-{description}
• Commit format: type(scope): description [BD-{number}]
• Tests must pass before marking complete
• Lead Dev reviews before merge

---

Configuration

# bangladev.yaml (created by `bangladev init`)
repo:
  root: /path/to/repo
  name: my-project

llm:
  endpoint: http://localhost:8000/v1
  model: minimax-m2.5
  fast_model: qwen3-30b-a3b  # optional, for QA/DevOps
  api_key: none
  max_tokens: 8192
  temperature: 0.1

team:
  lead:      { enabled: true, heartbeat_interval: 300 }
  architect: { enabled: true, heartbeat_interval: 600 }
  backend:   { enabled: true, heartbeat_interval: 120 }
  frontend:  { enabled: true, heartbeat_interval: 120 }
  qa:        { enabled: true, heartbeat_interval: 180 }
  devops:    { enabled: false, heartbeat_interval: 600 }

budgets:
  monthly_token_limit: 50000000
  per_agent_limit: 10000000

workspace:
  strategy: worktree
  branch_template: "bangladev/{agent}/{issue}"
  auto_pr: true

---

Architecture

bangladev/
├── packages/
│   ├── adapter-utils/          # Shared adapter types (from Paperclip)
│   ├── adapters/
│   │   ├── aider-local/        # Aider CLI wrapper for code editing
│   │   └── vllm-direct/        # Direct OpenAI-compatible chat
│   ├── db/                     # Drizzle/PostgreSQL schema + migrations
│   ├── repo-context/           # Repository scanner + context engine
│   └── shared/                 # Shared utilities
├── server/                     # Express API server (from Paperclip core)
├── ui/                         # React dashboard (from Paperclip core)
├── cli/                        # CLI commands (init, start, task, dashboard)
├── skills/                     # Agent skill definitions
│   ├── bangladev/              # Core heartbeat coordination
│   ├── git-ops/                # Branch, commit, PR operations
│   ├── test-runner/            # Language-aware test execution
│   ├── code-review/            # Review checklists and gates
│   ├── codebase-index/         # Codebase exploration procedures
│   └── deploy/                 # CI/CD and deployment
├── seed/                       # Dev team provisioning
└── STRIPPING-GUIDE.md          # Full Paperclip → BanglaDev migration map

---

Credits

BanglaDev is a fork of Paperclip AI (MIT License). The core heartbeat scheduler, task system, budget enforcement, and plugin architecture are retained from Paperclip. Everything else has been stripped, replaced, or rewritten for local-LLM, dev-only use.