opencode-delegate

OpenCode Delegate lets an orchestrating agent hand a bounded coding task to the separate OpenCode CLI as a background implementer, then review the resulting diff and land it. It solves the problem of parallelizing implementation work while keeping a human or orchestrator firmly in the reviewer/committer role: the orchestrator writes the brief and owns judgment, OpenCode does the typing in its own session, and the orchestrator verifies and commits. It is meant for tasks large enough to justify delegation but where the diff will be reviewed before commit.

The workflow is a five-step loop: write a self-contained brief (OpenCode sees only the brief plus the working tree, so goals, constraints, real gate commands, and a report contract all go in it), dispatch it via a bundled Node relay that wraps `opencode run`, poll for completion, review the diff, and land it. The skill stresses model selection (OpenCode has no default model; the human owns the allowed set, the orchestrator picks per task) and provides a detailed review discipline: check for weakened or skipped tests, re-run the project's actual gates rather than trusting the self-reported `result.json`, round-trip migrations, grep for dangling references after removals, and sweep generated code for hardcoded success values, catch-all error suppression, unverified imports, dead weight, and speculative surface. Rework is sent as a delta via `--resume-last`.

It targets engineers and agent orchestrators (primarily Claude Code) who want to delegate mechanical or well-scoped implementation work while staying the reviewer. Prerequisites include the authenticated `opencode` CLI, Node 18+, git, and shell access.