opencode-plugin-cc

Use Opencode from inside Claude Code: delegate coding tasks as background jobs, get second-model code reviews, and ask questions — without inflating your Claude Code context window.

Inspired by codex-plugin-cc, built on Opencode's richer CLI: resumable sessions, transcript export, multi-provider models (Anthropic, OpenAI, Google, Z.AI, local, …), and an optional headless server.

Why

Claude Code is great at orchestrating work — but every long task it runs inline burns its own context window. This plugin hands the heavy lifting to Opencode:

• Your session stays free. Delegation returns a job ID in three lines; Claude is notified automatically when the job finishes.
• Your context stays small. Results come back as a compact summary (status + session ID + log tail). Full transcripts stay on disk until you actually need them.
• A second model's perspective. Run reviews and implementations on a different model than the one orchestrating, on your own Opencode provider setup.

Commands

Command	Description
/opencode:setup	Verify Opencode is installed, authenticated, and ready
/opencode:ask	Ask Opencode a question (foreground, inline answer)
/opencode:review	Code review of pending changes (read-only; diff via temp file)
/opencode:delegate	Hand off a coding task as a background job with auto-notification
/opencode:status	List background jobs + recent Opencode sessions
/opencode:result	Fetch a job's result (status, session ID, log tail)
/opencode:cancel	Kill a running background job
/opencode:serve	Manage the optional headless server (start/stop/status)

Plus the opencode-worker subagent: delegates, waits, verifies the work (git diff, tests), and reports back compactly.

Requirements

• Claude Code with plugin support
• Opencode CLI ≥ 1.17 — curl -fsSL https://opencode.ai/install | bash
• A provider configured in Opencode — opencode auth login
• macOS or Linux (bash, git; jq optional but recommended)

Install

/plugin marketplace add brahmsyaifullah/opencode-plugin-cc
/plugin install opencode@opencode-marketplace

Then verify:

/opencode:setup

Usage

/opencode:ask How does this project's build system work?
/opencode:review main
/opencode:delegate Fix the failing tests in src/auth — run `npm test` to verify

A delegation looks like this:

> /opencode:delegate Refactor the database pool to use lazy connections

  JOB_ID: oc-20260713-231632-18274
  Task delegated to Opencode in background.

  I'll report back when it completes. (Claude's context stays free —
  the job's full transcript never enters the conversation.)

  … later, automatically …

  Opencode finished: pool.ts and config.ts modified, `npm test` passing.
  Session ses_0a3bb… available for follow-ups.

Follow-up work continues Opencode's own session (-s <session-id>), so nothing needs re-explaining.

Automatic Delegation (no slash command)

The worker agent and delegate skill carry proactive trigger descriptions, so orchestrator models can delegate on their own. The decision rule they encode:

Can the task be written as a self-contained spec — files to touch, expected behavior, and a verification command? If yes, delegate; if it needs iterative judgment or conversation context, keep it inline.

For near-deterministic routing, add the same rule to your CLAUDE.md:

## Opencode delegation policy
Before starting implementation work, ask: can I write a SELF-CONTAINED spec for it —
files to touch, expected behavior, and a verification command?
- YES → delegate to the opencode-worker agent instead of implementing inline.
- NO (needs iterative judgment or conversation context) → keep inline.
Keep orchestration, review, and verification in the main session.

LLM routing is never 100% guaranteed, but a binary yes/no rule like this gets very high compliance.

Context-Window Strategy

1. Delegate = background job. /opencode:delegate returns a JOB_ID immediately; a backgrounded wait notifies Claude when the job finishes. No polling, zero context cost while it runs.
2. Results come back compact. Status + session ID + log tail — not the full transcript.
3. Full history stays on disk. Job logs in ~/.claude/opencode-jobs/<job-id>/; complete transcripts via opencode export <session-id>, pulled into context only when needed.
4. Diffs never enter context. /opencode:review writes the diff to a temp file and attaches it with -f.
5. Follow-ups reuse Opencode's memory. Resume with -s <session-id> instead of re-explaining.

Architecture

Claude Code
  └─ /opencode:* commands, opencode-worker agent
       ├─ scripts/opencode-bridge.sh
       │    ├─ run      → opencode run --auto "prompt"           (foreground)
       │    ├─ delegate → background job → ~/.claude/opencode-jobs/<id>/
       │    │              (pid, prompt.txt, output.log, exit-code)
       │    ├─ review   → git diff → temp file → opencode run -f diff
       │    ├─ wait     → blocks until job done, prints result (notification hook)
       │    └─ status/result/cancel/gc → job dir + opencode session list/export
       └─ scripts/opencode-serve.sh (optional headless server)
            └─ start/stop/status → opencode serve on 127.0.0.1:4096

Each job's ID doubles as its Opencode session title — jobs map 1:1 to resumable sessions.

Server attach mode (optional)

With a headless server running (/opencode:serve start), every bridge call attaches automatically (opencode run --attach <url> --dir <cwd>): lighter per-call client, shared session state, one warm server instead of a cold CLI per job. No server → transparent fallback to standalone CLI. Both modes fully supported.

• Binds to 127.0.0.1 only; serves all projects (--dir pins each task to the caller's cwd).
• Caveat: cancelling in attach mode kills the local client; the server-side session may run to completion.

Configuration

Variable	Default	Description
OPENCODE_CMD	opencode	Path to the Opencode CLI
OPENCODE_MODEL	(Opencode default)	Model as provider/model; list with opencode models
OPENCODE_JOBS_DIR	~/.claude/opencode-jobs	Background job state
OPENCODE_RESULT_LINES	120	Log lines returned by result
OPENCODE_SERVER_PORT	4096	Headless server port
OPENCODE_ATTACH	(auto)	off = force standalone; <url> = attach to external server

Pin a default model for everything Opencode runs in ~/.config/opencode/opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "zai-coding-plan/glm-5.2"
}

Troubleshooting

Symptom	Fix
Opencode CLI not found	Install (see Requirements); the bridge already checks ~/.opencode/bin
Opencode >= 1.17 required	opencode upgrade
Job stuck running	/opencode:cancel <job-id>, inspect ~/.claude/opencode-jobs/<id>/output.log
Job failed(N)	/opencode:result <job-id> shows the error tail; auth/model issues surface here
Empty answers / auth errors	opencode auth login, then /opencode:setup
Old jobs piling up	Automatic GC removes finished jobs after 7 days; manual: opencode-bridge.sh gc [days]

Uninstall

/plugin uninstall opencode@opencode-marketplace
/plugin marketplace remove opencode-marketplace

Job state lives in ~/.claude/opencode-jobs/, server state in ~/.claude/opencode-server/ — delete them if you want a clean sweep. Opencode itself is untouched.

Contributing

Issues and PRs welcome at brahmsyaifullah/opencode-plugin-cc. Validate before submitting:

claude plugin validate .
bash -n scripts/*.sh

License

MIT