opencode-plugin-cc

Run opencode as a subagent inside Claude Code.

Get a second opinion on your diff, or hand off a long task to a different model — without leaving Claude Code.

A community port of openai/codex-plugin-cc that drives your local opencode CLI instead of Codex. Not affiliated with OpenAI, Anthropic, or the opencode project.

---

Why

Claude Code is good at writing code, and less good at grading its own homework. This plugin puts a second, independent model in the loop:

• Review — a fresh agent reads your diff with no memory of why you wrote it that way.
• Delegate — long or grindy tasks run in the background on a cheaper model while Claude stays responsive.
• Second opinion — different provider, different failure modes. Any provider opencode supports works, including the OpenCode Go plan (opencode-go/*).

Features

Command	What it does
/opencode:review	Read-only review of your working tree or branch diff
/opencode:adversarial-review	Same, but the prompt attacks the design: tradeoffs, assumptions, failure modes
/opencode:rescue	Hand a task to opencode (write-capable by default)
/opencode:status	List background jobs
/opencode:result	Print a finished job's output
/opencode:cancel	Kill a running job
/opencode:setup	Readiness check + toggle the optional stop-time review gate

Plus the opencode-rescue subagent, so plain English works too: "Ask opencode to redesign the DB connection to be more resilient."

Requirements

• Claude Code
• opencode CLI on PATH, signed in (opencode providers login)
• Node.js ≥ 18.18

No other dependencies. The plugin is ~400 lines of Node stdlib.

Install

From this marketplace:

/plugin marketplace add hieuvu7/opencode-plugin-cc

/plugin install opencode@opencode-cc

/reload-plugins

Then verify your setup:

/opencode:setup

Or from a local clone — point the marketplace at the checkout, and edits to the plugin apply without reinstalling:

git clone https://github.com/hieuvu7/opencode-plugin-cc.git

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

Usage

Reviews

/opencode:review                    # uncommitted working tree
/opencode:review --base main        # everything on this branch
/opencode:review --background       # returns a job id immediately

opencode's read-only plan agent inspects the diff with git itself, so nothing gets edited. Multi-file reviews take a while — prefer --background, then /opencode:status and /opencode:result.

Adversarial mode takes optional focus text after the flags:

/opencode:adversarial-review --base main challenge the caching and retry design
/opencode:adversarial-review --background look for race conditions

Delegating work

/opencode:rescue investigate why the tests started failing
/opencode:rescue --read-only dig into the regression
/opencode:rescue --model opencode-go/kimi-k3 --variant high fix the flaky integration test
/opencode:rescue --resume apply the top fix from the last run

• write-capable by default (opencode build agent); --read-only for investigation only
• omit --model / --variant and your own opencode config defaults apply
• models are provider/model — run opencode models to list them
• --resume continues the newest opencode thread from this repo; --fresh forces a new one
• a job is killed as stalled if opencode emits no output for 2 minutes after start or 10 minutes mid-run; --timeout-ms <n> overrides both windows

Managing jobs

/opencode:status
/opencode:result task-abc123
/opencode:cancel

/opencode:result prints the stored output plus an opencode --session <id> command, so you can pick up the same thread in the opencode TUI.

The review gate (optional)

/opencode:setup --enable-review-gate
/opencode:setup --disable-review-gate

With the gate on, a Stop hook runs a working-tree review whenever Claude tries to finish. A verdict of FIX FIRST blocks the stop and feeds the findings back so Claude addresses them.

Warning

The gate can create a long Claude↔opencode loop and burn quota. Only enable it while you are watching the session. It is off by default.

How it works

Claude Code
  └─ commands / opencode-rescue subagent
       └─ scripts/opencode-companion.mjs
            └─ opencode run --format json --auto --agent <plan|build> …

The companion spawns the CLI, folds its NDJSON event stream into a single result, and stores jobs (id, status, session id, output) under a per-workspace state directory. --background re-execs the companion detached and returns a job id right away. Your own opencode.json and agent config are used as-is.

Two details that matter for headless opencode, learned the hard way:

• stdin is closed — opencode run hangs forever on an inherited non-TTY stdin.
• --auto is passed — otherwise permission prompts block a run with nobody there to answer.

Differences from the Codex plugin

• No /transfer equivalent — opencode's import consumes its own export format, not Claude's .jsonl.
• One companion script over the plain CLI, instead of the Codex app-server JSON-RPC broker (~400 lines vs ~6k).

Development

npm test

Pure node:test, no framework. Contributions welcome — issues and PRs are the right place for bugs, new commands, and provider quirks. Keep the dependency count at zero.

Layout:

.claude-plugin/marketplace.json
plugins/opencode/
  .claude-plugin/plugin.json
  commands/       # 7 slash commands
  agents/         # opencode-rescue subagent
  hooks/          # optional Stop review gate
  scripts/        # opencode-companion.mjs + hook
tests/

License

Apache-2.0.