Opencode Builder

OpenCode Builder

A comprehensive skill for building on top of OpenCode — an AI-powered coding assistant. This skill covers both the JavaScript/TypeScript SDK for programmatic control and the plugin architecture for extending OpenCode's behavior.

Table of Contents

1. [Quick Start: Choose Your Path](#quick-start-choose-your-path)
2. [SDK Path: Programmatic Control](#sdk-path-programmatic-control)
3. [Plugin Path: Extending OpenCode](#plugin-path-extending-opencode)
4. [Configuration Deep-Dive](#configuration-deep-dive)
5. [Common Patterns](#common-patterns)
6. [Advanced Orchestration](#advanced-orchestration)
7. [Debugging & Troubleshooting](#debugging--troubleshooting)
8. [Reference Files](#reference-files)

---

Quick Start: Choose Your Path

OpenCode offers two primary extension surfaces:

| Path | Use When | Key Packages | |------|----------|-------------| | SDK (@opencode-ai/sdk) | You want to control OpenCode from an external app, script, or service. Build integrations, automate workflows, or create UIs that interact with OpenCode. | @opencode-ai/sdk | | Plugin (@opencode-ai/plugin) | You want to run code inside OpenCode, hook into its lifecycle, add custom AI tools, or modify behavior. | @opencode-ai/plugin |

You can use both in the same project. A plugin can internally use the SDK client to call back into OpenCode.

---

SDK Path: Programmatic Control

Installation

npm install @opencode-ai/sdk
# or
bun add @opencode-ai/sdk

Two Client Modes

Mode 1: Full Lifecycle — starts a local server + client

import { createOpencode } from "@opencode-ai/sdk"

const { client, server } = await createOpencode({
  hostname: "127.0.0.1",
  port: 4096,
  timeout: 5000,          // ms to wait for server start
  config: {               // overrides / merges with opencode.json
    model: "anthropic/claude-3-5-sonnet-20241022",
  },
})

// Use client...
await server.close()

Mode 2: Client Only — connects to an already-running OpenCode server

import { createOpencodeClient } from "@opencode-ai/sdk"

const client = createOpencodeClient({
  baseUrl: "http://localhost:4096",
  throwOnError: false,     // return error objects instead of throwing
  responseStyle: "fields", // or "data"
})

Core Workflows

1. Session Management

Sessions are the primary unit of interaction in OpenCode. Create them, send prompts, and read messages.

// Create a session
const session = await client.session.create({
  body: { title: "My automation task" },
})

// Initialize a session (analyzes project and creates AGENTS.md context)
await client.session.init({ path: { id: session.id } })

// Send a prompt and get AI response
const result = await client.session.prompt({
  path: { id: session.id },
  body: {
    parts: [{ type: "text", text: "Refactor the auth module" }],
  },
})
// result.data.info contains the assistant message metadata
// result.data.parts contains message parts (text, tool calls, etc.)

// Inject context WITHOUT triggering a response (great for setup/context)
await client.session.prompt({
  path: { id: session.id },
  body: {
    noReply: true,
    parts: [{ type: "text", text: "You are working in a Next.js project." }],
  },
})

// Run a shell command through the session
const shellResult = await client.session.shell({
  path: { id: session.id },
  body: { command: "npm test" },
})

// List all messages in a session
const messages = await client.session.messages({ path: { id: session.id } })

2. Structured Output

Request validated JSON from the model using JSON Schema:

const result = await client.session.prompt({
  path: { id: sessionId },
  body: {
    parts: [{ type: "text", text: "Analyze this codebase" }],
    format: {
      type: "json_schema",
      schema: {
        type: "object",
        properties: {
          summary: { type: "string" },
          techStack: { type: "array", items: { type: "string" } },
          entryPoints: { type: "array", items: { type: "string" } },
        },
        required: ["summary", "techStack"],
      },
      retryCount: 2,  // retries if validation fails (default: 2)
    },
  },
})

console.log(result.data.info.structured_output)
// { summary: "...", techStack: ["React", "TypeScript"], entryPoints: [...] }

// Check for structured output errors
if (result.data.info.error?.name === "StructuredOutputError") {
  console.error("Failed after", result.data.info.error.retries, "attempts")
}

Best practices for structured output:

• Provide clear description fields on schema properties — the model uses these
• Keep schemas focused; deeply nested schemas are harder to fill correctly
• Increase retryCount for complex schemas, decrease for simple ones
• Always handle StructuredOutputError gracefully

3. File Operations

// Read a file
const file = await client.file.read({ query: { path: "src/index.ts" } })
// file.data.content has the raw text
// file.data.type is "raw" or "patch"

// Search for text across files
const matches = await client.find.text({
  query: { pattern: "function.*handler" },
})
// matches.data is an array of { path, lines, line_number, submatches }

// Find files by name
const files = await client.find.files({
  query: { query: "*.ts", type: "file", limit: 50 },
})

// Find directories
const dirs = await client.find.files({
  query: { query: "components", type: "directory" },
})

// Get git status for tracked files
const status = await client.file.status()

4. TUI Control

Control the OpenCode terminal interface programmatically:

await client.tui.appendPrompt({ body: { text: "npm install lodash" } })
await client.tui.submitPrompt()          // Press "enter"
await client.tui.showToast({
  body: { message: "Done!", variant: "success" },
})
await client.tui.openHelp()
await client.tui.openSessions()
await client.tui.openModels()
await client.tui.executeCommand({ body: { command: "/clear" } })

5. Event Streaming

Subscribe to real-time events via Server-Sent Events:

const events = await client.event.subscribe()
for await (const event of events.stream) {
  console.log(event.type, event.properties)
  // event.type examples: "session.updated", "message.part.updated", "tool.execute.after"
}

6. Authentication

Set API keys for third-party providers that OpenCode will use:

await client.auth.set({
  path: { id: "anthropic" },
  body: { type: "api", key: "sk-ant-api03-..." },
})

7. Logging

Write structured log entries that appear in OpenCode's internal logs:

await client.app.log({
  body: {
    service: "my-integration",
    level: "info",    // debug | info | warn | error
    message: "Deployment started",
    extra: { env: "production", version: "1.2.3" },
  },
})

SDK Type Safety

All types are generated from OpenCode's OpenAPI spec:

import type { Session, Message, Part, Config, Project } from "@opencode-ai/sdk"

---

Plugin Path: Extending OpenCode

Installation

npm install @opencode-ai/plugin
# or
bun add @opencode-ai/plugin

Plugin Basics

A plugin is a JS/TS module that exports a function receiving a context object and returning hooks:

import type { Plugin } from "@opencode-ai/plugin"

export const MyPlugin: Plugin = async ({ project, client, $, directory, worktree }) => {
  // Initialization code runs here
  console.log("Plugin loaded for project:", project?.name)

  return {
    // Hook implementations go here
  }
}

Context object (ctx) properties:

| Property | Type | Description | |----------|------|-------------| | project | Project \| null | Current project metadata | | directory | string | Current working directory | | worktree | string | Git worktree path | | client | OpencodeClient | SDK client for calling back into OpenCode | | $ | BunShell | Bun's shell API for executing commands |

Loading Plugins

Plugins load from two sources:

1. Local files (auto-loaded at startup)

• .opencode/plugins/ — project-level
• ~/.config/opencode/plugins/ — global

2. npm packages (installed automatically via Bun)

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-helicone-session", "@my-org/custom-plugin"]
}

Load order: Global config → Project config → Global plugin dir → Project plugin dir

Dependencies for local plugins: Add a package.json to .opencode/:

{
  "dependencies": {
    "shescape": "^2.1.0"
  }
}

OpenCode runs bun install at startup.

Custom Tools

Add AI-accessible tools that OpenCode can invoke:

import { Plugin, tool } from "@opencode-ai/plugin"

export const CustomToolsPlugin: Plugin = async (ctx) => {
  return {
    tool: {
      deploy_to_vercel: tool({
        description: "Deploy the current project to Vercel",
        args: {
          preview: tool.schema.boolean().optional(),
          environment: tool.schema.enum(["production", "staging"]).optional(),
        },
        async execute(args, context) {
          const { directory, worktree } = context
          // args.preview, args.environment are typed and validated
          const { $ } = ctx
          const result = await $`cd ${directory} && vercel ${args.preview ? "--preview" : ""}`
          return `Deployed! Output: ${result.stdout}`
        },
      }),
    },
  }
}

Tool precedence: If a plugin tool shares a name with a built-in tool, the plugin's version wins.

Tool schema helpers:

• tool.schema.string()
• tool.schema.number()
• tool.schema.boolean()
• tool.schema.enum([...])
• tool.schema.array(itemSchema)
• .optional() — make any field optional
• .describe("help text") — add field descriptions

Lifecycle Hooks

Shell / Tool Interception

return {
  // Runs BEFORE any tool executes — modify args or block actions
  "tool.execute.before": async (input, output) => {
    if (input.tool === "read" && output.args.filePath.includes(".env")) {
      throw new Error("Blocked: do not read .env files")
    }
    // Modify arguments
    if (input.tool === "bash") {
      const { escape } = await import("shescape")
      output.args.command = escape(output.args.command)
    }
  },

  // Runs AFTER tool execution — inspect or modify results
  "tool.execute.after": async (input, output) => {
    console.log(`Tool ${input.tool} finished with:`, output.result)
  },

  // Inject environment variables into ALL shell execution
  "shell.env": async (input, output) => {
    output.env.MY_API_KEY = process.env.MY_API_KEY
    output.env.PROJECT_ROOT = input.cwd
  },
}

Session Hooks

return {
  "session.created": async ({ event }) => {
    console.log("New session:", event.properties.sessionId)
  },
  "session.updated": async ({ event }) => {
    // Session metadata changed
  },
  "session.idle": async ({ event }) => {
    // AI finished responding — great time for notifications
  },
  "session.error": async ({ event }) => {
    console.error("Session error:", event.properties.error)
  },
}

Message Hooks

return {
  "message.updated": async ({ event }) => {
    // A message was modified
  },
  "message.part.updated": async ({ event }) => {
    // Streaming update to a message part
  },
  "message.part.removed": async ({ event }) => {
    // A part was deleted (e.g., user removed a tool call)
  },
}

TUI Hooks

return {
  "tui.prompt.append": async ({ event }) => {
    // User added text to the prompt
  },
  "tui.toast.show": async ({ event }) => {
    // A toast notification was shown
  },
  "tui.command.execute": async ({ event }) => {
    // User executed a slash command
  },
}

Compaction Hooks (Experimental)

Customize how session context is summarized when it gets too long:

return {
  "experimental.session.compacting": async (input, output) => {
    // Inject additional context into the compaction prompt
    output.context.push(`
## Project State
- Currently working on: auth refactor
- Important decisions: using JWT, not sessions
- Active files: src/auth.ts, src/middleware.ts
`)

    // OR replace the entire compaction prompt:
    output.prompt = `
You are generating a continuation prompt for a multi-agent session.
Summarize:
1. Current task and status
2. Files being modified
3. Blockers or dependencies
4. Next steps
`
  },
}

When output.prompt is set, it replaces the default compaction prompt entirely.

Event Hook Pattern

Generic event listener for any event type:

return {
  event: async ({ event }) => {
    if (event.type === "session.idle") {
      await ctx.client.app.log({
        body: { service: "my-plugin", level: "info", message: "Session completed" },
      })
    }
  },
}

Logging from Plugins

Always use client.app.log() instead of console.log for structured, actionable logging:

await ctx.client.app.log({
  body: {
    service: "my-plugin",
    level: "info",
    message: "Plugin initialized",
    extra: { version: "1.0.0" },
  },
})

Levels: debug, info, warn, error

---

Configuration Deep-Dive

opencode.json Schema

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anthropic/claude-3-5-sonnet-20241022",
  "provider": {
    "anthropic": {
      "name": "Anthropic",
      "api": "anthropic",
      "url": "https://api.anthropic.com"
    },
    "openai": {
      "name": "OpenAI",
      "api": "openai",
      "url": "https://api.openai.com/v1"
    },
    "custom": {
      "name": "My Proxy",
      "api": "openai",
      "url": "http://localhost:18921/v1"
    }
  },
  "plugin": [
    "opencode-helicone-session",
    "file:///Users/me/my-plugin/src/index.ts"
  ],
  "mcp": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/projects"]
    }
  }
}

Runtime note: OpenCode runs on Bun, so plugins can use Bun-native APIs like Bun.file(), Bun.write(), Bun.hash(), and top-level await without bundling.

Config locations (precedence: top to bottom):

1. ~/.config/opencode/opencode.json — global defaults
2. ./opencode.json — project-specific overrides

Key fields:

• model — default model ID (format: provider/model-id)
• provider — custom API endpoints, useful for proxies or self-hosted models
• plugin — npm package names or file:// paths to local plugins
• mcp — Model Context Protocol server configurations

TypeScript Plugin tsconfig

{
  "compilerOptions": {
    "target": "ES2022",
    "module": "ESNext",
    "moduleResolution": "bundler",
    "strict": true,
    "esModuleInterop": true,
    "skipLibCheck": true,
    "forceConsistentCasingInFileNames": true,
    "declaration": true
  }
}

---

Common Patterns

Pattern 1: Plugin That Starts a Background Service

A plugin that auto-starts a proxy server if not already running:

import { Plugin } from "@opencode-ai/plugin"
import { spawn } from "node:child_process"

const PORT = 18921
const PROXY = `http://localhost:${PORT}`

async function startProxy(): Promise {
  try {
    const res = await fetch(`${PROXY}/health`, { signal: AbortSignal.timeout(1000) })
    if (res.ok) return
  } catch { /* not running */ }

  const child = spawn("node", ["proxy.js", String(PORT)], {
    stdio: "ignore",
    detached: true,
  })
  child.unref()

  // Wait for startup
  for (let i = 0; i  setTimeout(r, 500))
  }
}

export const ServicePlugin: Plugin = async (ctx) => {
  await startProxy()

  return {
    "shell.env": async (_input, output) => {
      output.env.MY_PROXY_URL = PROXY
    },
  }
}

Pattern 2: Notification on Session Completion

import { Plugin }

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [Timmy6942025](https://github.com/Timmy6942025)
- **Source:** [Timmy6942025/opencode-builder-skill](https://github.com/Timmy6942025/opencode-builder-skill)
- **License:** MIT

Install and usage instructions live in the source repository linked above.