Skip to content

beast-ofcourse/opencode-autonomous-team

Repository files navigation

🤖 OpenCode Autonomous Team

Ship production software without writing a single line of code.

17 specialized AI agents orchestrate, research, design, implement, test, review, harden, and ship your projects — autonomously.

npm version GitHub stars MIT license Built for OpenCode CI status

npm install opencode-autonomous-team

🚀 Get Started in 60 Seconds

Prerequisites: OpenCode + Node.js 18+ + an AI model provider (Anthropic, OpenAI, etc.)

# Install the scaffold
npm install opencode-autonomous-team

# Or clone the repo
git clone https://github.com/beast-ofcourse/opencode-autonomous-team.git my-project
cd my-project

# Launch OpenCode and go
opencode

Your first command:

/start-project Build a habit-tracking web app. Users can create habits,
  check them off daily, and see a streak. Should work on mobile browsers.
  Free tier only — no paid infra.

The orchestrator takes over from here. It researches, plans, gets your approval, then builds, tests, reviews, hardens, and ships — end to end, without you writing code.

Command What It Does
/start-project <goal> Full autopilot: research → requirements → architecture → plan → stop for your approval
/build Execute: implement → test → lint → fix → review → optimize → document → harden → ship
AUTOPILOT=true Skip the approval stop — go from goal to shipped in one shot
/status Check progress without triggering new work
/replan <change> Change scope mid-project without losing history

🧠 What This Actually Does

Most AI coding tools are a single agent with a text editor. They write code, call it done, and hallucinate confidently about things they didn't test.

This project replaces that with a virtual engineering department:

🧠 Orchestrator — the tech lead who never sleeps
  ├── 🔍 Researcher     — competitor analysis, library evaluation, best-practice mining
  ├── 📐 Planner        — requirements engineering, architecture design, task breakdown
  ├── 🎨 Frontend       — UI components, state management, accessibility, styling
  ├── ⚙️ Backend        — APIs, databases, auth, business logic, integrations
  ├── 🧪 Tester         — unit/integration/e2e tests, coverage, real pass/fail results
  ├── 📊 Performance    — profiling, bundle analysis, query optimization, caching
  ├── 🔒 Security       — threat modeling, dependency audits, authN/Z review
  ├── 📖 Docs Writer    — README, API docs, changelog, deployment guides
  ├── 👁️ Reviewer       — independent code review, production-readiness gate
  ├── 🛠️ Perfectionist  — production hardening, fix tracking, 2-cycle audit gauntlet
  ├── 🐛 SWE Debugger   — reproduction-first debugging, root-cause analysis, minimal fixes
  ├── 🔬 SWE Testing    — test infrastructure, TDD workflows, property-based testing
  ├── 🧹 SWE Refactor   — behavior-preserving refactoring, dead code removal
  ├── 🚀 DevOps         — CI/CD pipelines, Docker, IaC, deployment strategies
  ├── 🛡️ SWE Security   — vulnerability remediation, dependency hardening
  └── ♿ UX Designer    — accessibility audits, WCAG compliance, design systems

Each agent is a specialist with scoped permissions. The frontend agent can't touch your database schema. The security agent can read everything but can't edit code — its judgment stays independent. No agent can spawn sub-agents (that's the orchestrator's job only), which means no runaway token chains.


📐 What Makes This Different

vs Vanilla AI Coding

Most AI Tools OpenCode Autonomous Team
One agent doing everything 17 specialists, each in its own lane
"I wrote code" = done 10-phase SDLC: research → requirements → architecture → plan → implement → test → review → optimize → document → ship
May or may not test 7 quality gates (A–G) with verified evidence before anything is marked done
Unlimited delegation ($$$) Depth-1 delegation — subagents can't spawn subagents. Token runaway engineered out
Security as an afterthought Security is a baseline — server-side validation always, passwords always hashed, UI always operable
One-shot prompt, no trail Living documents with revision logs — nothing gets silently rewritten

vs oh-my-openagent

This project and oh-my-openagent (65.6K ★, 3.3M+ downloads) are the two leading orchestration scaffolds for OpenCode. Here's an honest comparison:

Capability oh-my-openagent OpenCode Autonomous Team
Agent depth Flat orchestration 17 specialists with scoped permissions
SDLC rigor Task-based 10-phase lifecycle + 7 quality gates with evidence
Intent routing Built-in IntentGate Plugin-based intent_classify
Model routing Built-in Category-based (8 tiers → optimal model per task)
Background work Built-in parallel dispatch Plugin-based dispatch_background + dispatch_result
Checkpointing boulder.json Disk-persisted atomic checkpointing
Error recovery Standard 5-layer: retry → circuit breaker → fallback → degrade → fail
AI slop detection Not present tool.execute.after hook + clean_comments tool
API contracts Optional Mandatory before frontend/backend split
Team Mode viz ✅ Shipped Stub — on roadmap
Multi-harness ✅ Codex CLI edition On roadmap
Security Standard Hard-coded: no recursive delegation, no destructive commands

Choose oh-my-openagent if you want a mature ecosystem with Team Mode visualization and multi-platform support. Choose this project if you want rigorous SDLC process, evidence-gated quality, and depth-1 token safety.


🏗️ Architecture

                                 User
                                   │
                                   ▼
                    ┌─────────────────────────┐
                    │ Autonomous Orchestrator │
                    └────────────┬────────────┘
                                 │
              ┌──────────────────┼──────────────────┐
              │                  │                  │
              ▼                  ▼                  ▼
        Planning Layer     Execution Layer    Validation + Hardening Layer
      ┌──────────────┐   ┌──────────────┐   ┌──────────────────┐
      │ Research     │   │ Frontend     │   │ Testing          │
      │ Planning     │   │ Backend      │   │ Review           │
      │ Architecture │   │ Refactoring  │   │ Security         │
      └──────────────┘   └──────────────┘   │ Performance      │
                                             │ Perfectionist    │
                                             └──────────────────┘
                                  │
                                  ▼
                    ┌─────────────────────────┐
                    │   Tooling & Commands    │
                    │ Git • Terminal • Docs   │
                    │ Browser • MCP • CI/CD   │
                    └────────────┬────────────┘
                                  │
                                  ▼
                          Project Workspace

Depth-1 delegation is enforced at the permission layer, not by convention. Every subagent has task: deny — only the orchestrator can dispatch work. This is a hard architectural constraint, not a guideline. It prevents the uncontrolled recursive delegation that burns tokens and produces diminishing returns in other agent systems.


🔌 Plugin System

The team runs as a custom OpenCode plugin (team-plugin.ts) with 7 tools and 3 event hooks:

Custom Tools

Tool What It Does
team_status Report phase, agents, version, and MCP config
dispatch_background Spawn a child subagent session — returns dispatch_id immediately
dispatch_result Retrieve completed dispatch output from background agents
get_dispatch Check a dispatch's status without blocking
list_dispatches List all dispatches with optional status filter
intent_classify Classify user input into 6 intents (research/implement/investigate/fix/evaluate/open_ended)
clean_comments Scan and remove AI slop comment patterns from files

Intent Gate

Every user request runs through intent_classify before the orchestrator dispatches:

  • researchresearcher
  • implementfrontend / backend
  • investigateswe-debugger
  • fixswe-debugger (root cause) → frontend/backend (fix)
  • evaluatereviewer
  • open_ended → Phase 0-1 research first

Low-confidence classifications fall back to open_ended, triggering research before any code is written — preventing the "start coding the wrong thing" trap.

Checkpointing

State is persisted atomically to .opencode/plugins/.checkpoint.json on every phase change and session compaction. On restart, the team picks up exactly where it left off — phase, completed tasks, blocked items, and active dispatches are all restored. Schema includes a version field for future migrations.

AI Slop Detection

A tool.execute.after hook scans every file written by edit tools against 10 regex patterns ("Certainly!", "I'll ", "Let me ", "As an AI", etc.). Matches are logged as warnings (non-blocking). The clean_comments tool provides bulk scan-and-strip with a dry-run mode.


👥 Team Reference

Agent Role
orchestrator Primary agent — owns full goal-to-production lifecycle
researcher Competitor analysis, OSS prior art, library comparisons, best practices
planner Requirements engineering, architecture design, task breakdown
frontend UI components, state management, accessibility, styling
backend APIs, DB schema/migrations, auth, server logic, integrations
tester Unit, integration, contract, and e2e tests; coverage; fixtures
performance Profiling, bundle analysis, query optimization, caching, load-testing
security Threat modeling, dependency audits, authN/Z review, secrets hygiene
docs-writer README, API docs, architecture docs, changelog, deployment guides
reviewer Independent code review, production-readiness gate
perfectionist Production hardening — fixes findings from security + reviewer
swe-debugger Reproduction-first debugging, root-cause analysis, minimal fixes
swe-testing Test infrastructure, TDD workflows, property-based testing
swe-refactor Behavior-preserving refactoring, dead code removal
devops CI/CD, Docker, IaC, deployment, secrets management
swe-security Vulnerability remediation, dependency patching, secure config
ux-designer Accessibility audits, WCAG compliance, design system review

⚙️ Configuration

Models

Edit opencode.json at your project root:

{
  "model": "anthropic/claude-sonnet-4-5",
  "small_model": "anthropic/claude-haiku-4-5"
}

Category routing maps task complexity to optimal models automatically:

Category Model For
quick Haiku One-file fixes, config changes
unspecified-low Haiku Small, well-defined tasks
unspecified-high Sonnet Multi-file, moderate complexity
deep Opus Autonomous research + end-to-end
ultrabrain Opus Hard logic, algorithms, architecture
visual-engineering Sonnet UI, styling, animation
writing Haiku Documentation, changelogs
artistry Opus Creative, unconventional

One-Shot Mode

# Plan + build + deploy — no stops
AUTOPILOT=true AUTODEPLOY=true /start-project Build a habit tracker...

Without these env vars, the team stops after Phase 5 (planning) for your review before writing any code.

Adding Specialists

  1. Create .opencode/agents/<name>.md following the existing agent pattern
  2. Add "<name>": "allow" to the orchestrator's permission.task block
  3. Add a row in the orchestrator's specialist table
  4. Restart OpenCode

🔒 Safety Architecture

Constraint How It's Enforced
No recursive delegation Every subagent has task: deny. Orchestrator only, depth-1 max.
No destructive commands rm -rf, sudo, force-pushes, DB drops denied at permission layer. No agent can bypass this.
No fabricated results tester, performance, security report only what they actually ran.
No silent scope changes Every living doc has a revision log.
No rubber-stamp reviews reviewer is read-only. Its judgment stays independent.

🧩 Use Cases

  • Greenfield projects — from idea to shipped MVP, fully autonomous
  • Existing codebases — drop the team anywhere; it detects conventions and works within them
  • Prototype validation — working, tested prototype in hours, not days
  • Technical spikes — delegate research and proof-of-concept work to the team
  • Learning accelerator — watch the team design and build; study the architecture docs and test strategies it produces

📊 Project Status

Production-ready and actively maintained. The agent definitions, permission model, living docs system, and plugin architecture have been hardened through real use across multiple project types. 16 specialist agents + orchestrator, 7 custom plugin tools, 3 event hooks, disk-persisted checkpointing, category-based model routing, and a 10-phase SDLC with 7 quality gates.


🤝 Contributing

Contributions welcome! Open an issue or pull request on GitHub.


📄 License

MIT © Beast Ofcourse


Built for the OpenCode ecosystem.

About

A autonomous engineering team for OpenCode — implementing a full goal-to-production software development lifecycle.

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages