opencode-delegate

A Claude Code plugin that runs opencode-driven development: Claude
brainstorms, writes the plan, and writes every brief; opencode
sessions on free models — Muse Spark 1.3 by default — implement each task,
review each task, and review the whole branch. It is a mirror of
superpowers' subagent-driven-development with opencode in every subagent
seat.

Claude thinks and adjudicates. opencode models do the engineering. You
approve the plan once.

This repo is also a Claude Code plugin marketplace, so it installs on any
machine and works in every project.

How it works

1. You ask for something worth a plan — a feature, a refactor, a test
   suite. Claude invokes the skill itself (or you type
   /opencode-driven-development).
2. Claude brainstorms with you until the design is settled, writes an
   implementation plan (file map, bite-sized tasks with the code and tests
   spelled out), and shows it to you. You approve the plan once.
3. For each task, Claude writes a brief and runs oc-task. That creates a
   git worktree on branch oc/<plan> (first task) and dispatches a fresh
   opencode session per task inside it, permissions auto-approved because
   the worktree is the sandbox. The model implements, commits per step, runs
   the verification command, and ends with a ## RESULT block.
4. Claude re-runs the verification command itself, then dispatches an
   opencode reviewer (--role review, opencode's read-only plan agent)
   with the task's diff. Findings go into a fix loop — resume the same
   session up to three times, then a fresh session on a different free
   model — with a scoped re-review after every fix. A ledger in .oc-runs/
   records every decision.
5. After the last task, a final whole-branch review on the strongest free
   model. Then Claude reports every ruling it made and offers: merge now
   (oc-task --merge, --no-ff, one revert point via oc-undo), open a
   PR, or leave the branch.

Nothing stops for you between plan approval and the end except the four
things that should: a destructive operation, a security-sensitive action, a
side effect outside the worktree (merge/push/publish), or a plan so broken
every path is a guess. Tasks touching auth, secrets, crypto, or payments
get one extra question: whether you're OK sending that code to an opencode
model.

Prerequis