OpenCode

OpenCode is Shiba’s default coding agent harness. It is open source, highly extensible, and supports multi-provider LLM backends through your Cloudflare AI Gateway.

Agent Overview

Attribute	Specification
Harness ID	`opencode` (default)
CLI Binary	`opencode`
Pinned Version	`1.18.31` (in Dockerfile)
Default Model	`google/gemini-3.5-flash-lite`
Supported Providers	Google AI Studio, Anthropic, OpenAI, xAI (Grok), OpenCode Go
Credential Model	AI Gateway BYOK (dummy key inside sandbox; real key injected at egress)
Invocation	`opencode run "<task>" --format json`

Multi-Provider Support

OpenCode can route tasks across multiple AI providers by prefixing the model with the provider slug:

# Google Gemini
google/gemini-3.5-flash-lite
google/gemini-1.5-pro

# xAI Grok
xai/grok-4

# Anthropic Claude
anthropic/claude-sonnet-4-6

# OpenAI
openai/gpt-4o

How It Works

Developer Request
  └── CodingOrchestrator generates delegate_coding_task({ harness: "opencode" })
      └── Human signs off on exact prompt & repository
          └── Ephemeral Cloudflare Sandbox boots
              ├── Writes per-run .opencode.json configuration
              ├── Executes: opencode run "<task>" --format json
              ├── Streams newline-delimited JSON events
              └── Collects git status and builds unified diff

Configuration & Deployment

1. Configure Providers in AI Gateway

Add your provider keys (Google, Anthropic, OpenAI, xAI) to your Cloudflare AI Gateway.

2. Set Default Model

In wrangler.jsonc or via environment variables:

pnpm wrangler secret put CODING_MODEL # enter "google/gemini-3.5-flash-lite"

Selecting OpenCode in a Task

OpenCode is used by default if no harness is explicitly provided:

{
  "repoUrl": "https://github.com/acme/web",
  "task": "Add dark mode toggle to navigation bar",
  "baseBranch": "main",
  "harness": "opencode",
  "codingModel": "xai/grok-4"
}