opencode

opencode-plugin-cc

A Claude Code plugin that lets you delegate bounded, mechanical coding tasks from Claude (Opus) to opencode's cheaper secondary models — Kimi, Qwen, GLM, etc.

Opus stays as the orchestrator. Opencode does the grunt work. The long tool log stays inside the subagent's transcript, so orchestration doesn't burn your main Opus context.

Why use it

• Save Opus tokens. Grunt work runs on a cheap opencode model; only the final result lands in your main context.
• Explicit-only. The subagent fires on /opencode:delegate (or when you name opencode in the request), never as a surprise for generic tasks.

Requirements

• opencode ≥ 1.4.10 on PATH — brew install sst/tap/opencode or npm i -g opencode-ai
• Node.js ≥ 18.18
• At least one opencode provider authenticated (opencode auth login)

Install

/plugin marketplace add turygo/opencode-plugin-cc
/plugin install opencode@opencode-plugin-cc
/reload-plugins
/opencode:setup

Usage

Delegate a task to opencode:

/opencode:delegate add JSDoc to every exported function in src/auth/*.ts

Opus routes the request through the opencode-delegate subagent, which runs opencode run synchronously and returns the final output. Model selection follows your opencode config by default; pass --model <provider/model> to override per-call.

Run long tasks in the background:

/opencode:delegate --background port the validator test suite to vitest
# → [opencode-job] id=job-abc123 status=started
/opencode:status
/opencode:result job-abc123

Commands

Command	What it does
/opencode:delegate <task>	Hand a task to opencode. Blocks until done, or returns a job-id with --background.
/opencode:status [job-id]	List or inspect background jobs.
/opencode:result <job-id>	Print the stored output of a finished job.
/opencode:cancel [job-id]	Kill an active background job.
/opencode:setup	Verify opencode is installed and authenticated.

Flags for /opencode:delegate:

Flag	Purpose
--background / --wait	Run detached and return a job-id, or block until done (default --wait).
--model <provider/model>	Override opencode's default model for this call.
--continue	Continue opencode's last session instead of starting fresh.
--session <id>	Resume a specific opencode session.
--agent <name>	Run under a named opencode agent.
--dir <path>	Run opencode with a different working directory.

Everything after the flags becomes the opencode prompt.

Security

The companion always passes --dangerously-skip-permissions to opencode run — the spawned opencode has no TTY, so TUI permission prompts can never be answered. In practice opencode has full write and shell-exec authority in the current working directory while a delegation runs.

Always delegate inside a git repository and review the diff before committing.

License

MIT