opencode plugin for Claude Code

Use opencode from inside Claude Code for code reviews or to delegate tasks — a port of the official Codex plugin for Claude Code to the opencode CLI.

Because opencode fronts many providers (Anthropic, OpenAI, Google, open-weight and free opencode zen models), this effectively lets Claude Code hand work to whichever second model you have configured in opencode.

Project write-up: han-dong.link/en/work/opencode-plugin-cc

What You Get

• /opencode:review — a read-only opencode code review of your local git changes, with structured findings
• /opencode:adversarial-review — a steerable challenge review that attacks the design and assumptions
• /opencode:rescue — delegate investigation, debugging, or a full implementation task to opencode (write-capable by default, resumable)
• /opencode:transfer — hand this Claude Code session off into a resumable opencode session
• /opencode:status, /opencode:result, /opencode:cancel — manage background jobs
• /opencode:setup — readiness check, plus an optional stop-time review gate where opencode reviews every Claude turn that edited code

Requirements

• opencode CLI ≥ 1.x on your PATH (opencode --version to check). Tested against 1.17.x.
• A provider configured in opencode (opencode auth login). Free opencode zen models work without credentials; runs use whatever provider/model your opencode is set up for and count against that provider's usage.
• Node.js 20 or later

Install

Add the marketplace in Claude Code:

/plugin marketplace add handong66/opencode-plugin-cc

or from a local checkout:

/plugin marketplace add /path/to/opencode-plugin-cc

Install the plugin:

/plugin install opencode@opencode-plugin-cc

Reload plugins (/reload-plugins), then run:

/opencode:setup

/opencode:setup tells you whether the opencode CLI is ready. If opencode is missing, install it yourself with one of:

npm i -g opencode-ai
curl -fsSL https://opencode.ai/install | bash

If opencode has no stored credentials yet, run:

!opencode auth login

Usage

# Read-only review of uncommitted changes (structured verdict + findings)
/opencode:review

# Review the branch against a base ref
/opencode:review --base main

# Challenge the design; extra text steers the focus
/opencode:adversarial-review is the retry logic safe under concurrent writers?

# Delegate a task (write-capable by default; picks your opencode default model)
/opencode:rescue figure out why the login test is flaky and fix it

# Choose a model or reasoning variant explicitly
/opencode:rescue --model anthropic/claude-sonnet-4-5 --variant high refactor the cache layer

# Continue the previous opencode session
/opencode:rescue --resume apply the top fix

# Background jobs
/opencode:rescue --background port the parser to TypeScript
/opencode:status
/opencode:result
/opencode:cancel

# Hand this session off to opencode (costs one opencode model turn)
/opencode:transfer

Every finished run prints the opencode session id; continue it directly in opencode with opencode -s <session-id>.

Stop-time review gate

/opencode:setup --enable-review-gate makes opencode review every Claude turn that edited code before Claude is allowed to stop, blocking with concrete findings when something still needs fixing. It runs a full opencode turn on every stop — enable it only while actively monitoring a session, and disable it with /opencode:setup --disable-review-gate.

How it works

All commands go through one helper runtime, plugins/opencode/scripts/opencode-companion.mjs, which wraps headless opencode run --format json:

• Jobs: every run is tracked in per-workspace state (under Claude's plugin-data dir), so status/result/cancel work across foreground, background, and even across Claude sessions. Session-end hooks terminate any still-running jobs.
• Read-only vs write: read-only runs (reviews, plain rescue diagnosis) use opencode's built-in plan agent, which cannot edit files. Write-capable runs (task --write, the rescue default) pass --auto so opencode can act without interactive permission prompts.
• Structured reviews: the review prompts embed a JSON schema contract; the companion parses the model's final answer into verdict/findings/next-steps, falling back to raw output when a model refuses to cooperate.
• Transfer: opencode cannot import Claude transcripts natively, so transfer distills the Claude session transcript into a handoff prompt and seeds a fresh opencode session with it.

Development

npm test

Tests use node:test and a fake opencode fixture on PATH — no real model calls, no credentials needed.

License

Apache-2.0. This project is a derivative work modeled on the Codex plugin for Claude Code (Copyright OpenAI, Apache-2.0); see NOTICE. Not affiliated with Anomaly Innovations (opencode) or OpenAI.