This project enables delegating coding tasks to AI agents like OpenAI Codex and OpenCode, streamlining workflows for engineers and power users. It automates repetitive coding tasks (refactoring, migrations) while ensuring human oversight. The core benefit is a controlled, iterative development process where AI assists implementation, but the human reviewer retains final commit authority.

How It Works

It implements a "brief-dispatch-poll-review-land" loop. Users write self-contained briefs, dispatched via relay.mjs to AI CLIs (codex-delegate, opencode-delegate). After AI completion (returning result.json and diff), the user reviews, re-runs gates, and manually commits. This ensures the orchestrator, not the AI, owns the commit, offering a distinct advantage over automated systems.

Quick Start & Requirements

Install via Skills CLI: npx skills add amElnagdy/delegate-skills. Specific skills (codex-delegate, opencode-delegate) can be installed individually. Prerequisites include the respective AI CLIs (codex, opencode) installed and authenticated, Node.js 18+, and Git. An orchestrating agent capable of shell commands and file reading is required; examples assume bash/zsh.

Highlighted Details

• Integrates with the Skills CLI ecosystem.
• Offers codex-delegate (OpenAI Codex) and opencode-delegate (OpenCode CLI).
• Features inspectable relay.mjs scripts: no network calls, no credential access, no telemetry; only interacts with implementer CLI and Git.
• Supports multi-task queues for batch processing.
• opencode-delegate includes autonomy modes (build/plan) and direct brief piping.

Maintenance & Community

The provided README does not detail specific contributors, sponsorships, community channels, or a public roadmap.

Licensing & Compatibility

Licensed under MIT, permitting broad use, modification, and distribution, including commercial purposes. Compatibility with closed-source projects is generally assured, though users should note the terms of service for the underlying AI CLIs.

Limitations & Caveats

End-to-end verification of the delegate-review-commit loop is limited to Claude Code and remains formally unverified. OpenCode's cold start performance in constrained shells may require running tasks in standard terminals. Integration with other orchestrators (e.g., Cursor) is designed-for but unproven. A planned gemini-delegate skill is not yet implemented.