delegate-to-opencode

A Claude Code skill that delegates tasks to opencode — the open-source AI coding agent — non-interactively, then verifies the result and retries until the task is genuinely complete.

What it does

When you ask Claude Code to do something "using opencode" / "pakai opencode", this skill:

1. Translates your request into a self-contained opencode run call
2. Executes it with --dangerously-skip-permissions so it never hangs waiting for approval
3. Verifies the result — reads changed files, diffs, or test output to confirm the task is actually done
4. Retries via --continue if the result is incomplete or wrong, with a specific explanation of what's still missing
5. Reports what changed and which files were affected

Requirements

• Claude Code installed
• opencode installed and configured with a provider/API key (opencode providers)

Installation

Option A — copy the skill folder

cp -r . ~/.claude/skills/delegate-to-opencode

Option B — clone directly into your skills folder

git clone https://github.com/kabesma/delegate-to-opencode ~/.claude/skills/delegate-to-opencode

Restart Claude Code (or start a new session) for the skill to appear in the available skills list.

Usage

Just mention opencode as the executor in your request:

suruh opencode bikin fungsi validasi email di utils
have opencode fix the failing tests
lewat opencode, refactor the auth middleware
tell opencode to explain the auth flow in this repo
lanjutkan task opencode tadi, tambahin unit testnya

Claude Code will invoke this skill automatically — no slash command needed.

Verified user preferences (pre-configured)

Setting	Value	Reason
--dangerously-skip-permissions	always on	Prevents the bash process from hanging while waiting for approval
Confirmation before run	off	Run immediately once the task is clear
Default working directory	context-aware, falls back to project root	Runs in the directory most relevant to the task

How the retry loop works

opencode run "task" → verify result
        ↓ incomplete/wrong
opencode run --continue "here's what's still wrong: ..."
        ↓ verify again
repeat (max ~3 rounds) → report to user

Claude reads the actual file changes or output — not just opencode's "done" message — before deciding whether to retry.