OpenCode

Overview

OpenCode is an open-source AI coding agent that integrates with OpenClaw via ACP (Agent Client Protocol). It provides specialized coding capabilities including game development, multi-file projects, complex refactoring, and production-ready code generation.

Status: ✅ Fully tested - Built 12 production-ready games including Pac-Man, Space Invaders, Tetris, Q*bert, and more.

When to Use This Skill

Use OpenCode when:

Building games (HTML5 Canvas games work excellently)
Creating multi-file projects from scratch
Performing complex code refactoring across multiple files
Building production-ready applications with proper structure
Tasks requiring dedicated coding agent focus
The user explicitly asks to use OpenCode

Quick Start

Spawn OpenCode as ACP Subagent

// One-shot task (recommended for most builds)
sessions_spawn({
  runtime: "acp",
  agentId: "opencode",
  task: "Create a complete [game/app] in HTML/CSS/JavaScript...",
  mode: "run",
  cwd: "/path/to/project"
})

// Parallel builds (run multiple simultaneously)
sessions_spawn({...}) // Game 1
sessions_spawn({...}) // Game 2  
sessions_spawn({...}) // Game 3

Parallel Builds

OpenCode supports parallel execution - spawn multiple subagents simultaneously:

// Launch 3 games at once
const game1 = sessions_spawn({... Asteroids ...})
const game2 = sessions_spawn({... Doodle Jump ...})
const game3 = sessions_spawn({... Flappy Bird ...})

// All build concurrently, limited by slowest

Performance: 3 parallel games completed in ~1m 20s (same as single game)

Performance Expectations

Based on 12-game test suite:

Complexity	Examples	Build Time	Lines
------------	----------	------------	-------
Simple	Pong, Flappy Bird	30-60s	300-600
Medium	Snake, 2048, Breakout	1m - 1m 30s	500-800
Complex	Pac-Man, Q*bert, Tetris	1m 30s - 2m+	1000-1300

Proven Game Templates

The following game types have been fully tested and verified:

Game	Key Features	Test Result
------	--------------	-------------
Tetris	Ghost piece, hold, preview, levels	✅ Complete
Pac-Man	4 AI ghosts, maze, power pellets	✅ Complete
Space Invaders	Enemy formation, bunkers, UFO	✅ Complete
Snake	Grid movement, progressive speed	✅ Complete
Pong	AI opponent, physics	✅ Complete
Breakout	Multi-level, power-ups	✅ Complete
2048	Merge algorithm, undo, swipe	✅ Complete
Minesweeper	Flood-fill, 3 difficulties	✅ Complete
*Qbert**	Isometric, pyramid, enemies	✅ Complete (fixed spawn)
Asteroids	Vector physics, screen wrap	✅ Complete
Flappy Bird	Gravity, pipes, medals	✅ Complete
Doodle Jump	Bouncing, power-ups, black holes	✅ Complete (fixed spawn)

Commands Reference

Core Commands

Command	Description
---------	-------------
`opencode`	Start interactive TUI
`opencode run`	Run non-interactive mode
`opencode acp`	Start ACP server (for OpenClaw integration)
`opencode serve`	Start headless server
`opencode web`	Start web interface

Utility Commands

Command	Description
---------	-------------
`opencode providers`	Manage AI providers
`opencode models`	List available models
`opencode stats`	Show token usage
`opencode upgrade`	Update to latest version
`opencode session`	Manage sessions

Built-in Agents

build - Full-access agent for development work (default)
plan - Read-only agent for analysis and exploration
@general - Subagent for complex searches

Configuration

ACP Integration

OpenCode is configured in OpenClaw. The ACP command is:

opencode acp

This starts JSON-RPC over stdio for ACP-compatible communication.

Providers

OpenCode supports multiple AI providers:

OpenCode Zen (recommended)
Claude (Anthropic)
OpenAI
Google Gemini
Local models

Set API keys via opencode providers or environment variables.

Usage Patterns

Pattern 1: One-shot Build

sessions_spawn({
  runtime: "acp",
  agentId: "opencode",
  task: "Create a complete Snake game in HTML/CSS/JavaScript...",
  mode: "run",
  cwd: "/Users/jadegaul/.openclaw/workspace/codeprojects"
})

Pattern 2: Parallel Builds

// Build 3 games simultaneously
const builds = [
  sessions_spawn({ task: "Asteroids...", ... }),
  sessions_spawn({ task: "Doodle Jump...", ... }),
  sessions_spawn({ task: "Flappy Bird...", ... })
]

Pattern 3: Bug Fixes

// Fix existing code
sessions_spawn({
  runtime: "acp",
  agentId: "opencode",
  task: "Fix the bug in /path/to/file.html where...",
  mode: "run"
})

Pattern 4: Check Status

subagents({ action: "list" })

Pattern 5: Steer Running Subagent

subagents({
  action: "steer",
  target: "<runId>",
  message: "Continue with the implementation"
})

File Operations

OpenCode has full file system access:

Read/write files
Execute shell commands
Navigate directories
Git operations

Capabilities

100% open source (unlike Claude Code)
Provider-agnostic - use any model
LSP support - code intelligence
Client/server architecture - run remotely
Undo/redo - /undo and /redo commands
Session sharing - /share creates shareable links

Best Practices

Be specific in task descriptions - include file paths, required features
Use mode: "run" for one-shot builds (faster)
Use mode: "session" for multi-step work
Test immediately - some games need spawn fixes
Parallel builds work great for 3+ independent tasks

Common Issues & Solutions

Issue: Game spawns in wrong position

Fix: Spawn OpenCode to add starting platform/position

Issue: Character dies immediately

Fix: Add spawn delay or ensure safe starting position

Issue: Controls not working

Fix: Check that touch + keyboard controls both specified

References

See references/opencode-docs.md for complete ACP documentation.

Test Results Summary

Total Games Built: 12

Total Build Time: ~15 minutes

Total Lines of Code: ~8,000+

Success Rate: 100% (with minor spawn fixes)

Average Build Times:

Simple games: ~50s
Medium games: ~1m 15s
Complex games: ~1m 40s

OpenCode via ACP is production-ready for game development and coding tasks.

OpenCode Game Builder

概述