Build a Fleet of Coding-Agent Lanes with delegate-skills

Your orchestrator is good at planning and reviewing, but every bounded task it implements itself costs context. delegate-skills, a 2k-star MIT package by amElnagdy, hands those tasks to a separate coding CLI already installed on your machine — and keeps the review and the commit with you.

Why This Skill Matters

The package has one setup skill and eighteen implementer skills. delegate-setup discovers which implementer CLIs are installed, proposes a compact fleet of named lanes, shows you the complete configuration, and writes it — globally or per repository — only after explicit approval. It never dispatches work. Each lane binds a kind of work to one implementer plus optional dials such as model, effort or variant; once configured, you dispatch with the matching *-delegate skill and --lane <name>.

The eighteen *-delegate skills cover Aider, Google Antigravity, Claude Code, Cline, OpenAI Codex, Command Code, Cursor Agent, Grok Build, Kimi Code, Muse Code, OpenCode, Pi, Oh My Pi, Qoder, Mistral Vibe, GitHub Copilot CLI, Warp Agent CLI and Z.AI ZCode. Every relay speaks the same delegate-relay.result.v1 contract — status, exitCode, signal, the implementer's own final report, touchedFiles and a session id where the CLI exposes one — so you learn the loop once and swap the implementer freely.

Installation

npx skills add amElnagdy/delegate-skills --list
npx skills add amElnagdy/delegate-skills
npx skills add amElnagdy/delegate-skills --skill delegate-setup
npx skills add amElnagdy/delegate-skills --skill codex-delegate

--list browses first; the other three install the whole package, only the setup skill, or one implementer. --agent claude-code targets a specific agent and --global installs globally. To pin, append a release tag as @vMAJOR.MINOR.PATCH — the Skills CLI installs by git ref, not by metadata.version in SKILL.md. Requirements: Node 18+ and git, an orchestrating agent that can run shell commands and read files, and, for each *-delegate skill, its CLI authenticated as you would at the terminal.

Real Workflow: Create a Fleet, Then Delegate a Refactor

1. Ask your orchestrator to build the fleet:

Use $delegate-setup to discover my installed implementer CLIs and create a fleet for feature, tests, and UI work.

Setup discovers, proposes, shows the full config, and writes after you approve. The README's example fleet routes feature to OpenCode, tests to Codex and ui to Cursor.

2. Dispatch by lane, or skip setup and pick one implementer directly:

Use $opencode-delegate with --lane feature to implement the billing workflow, then review and commit it.
Use $codex-delegate to have Codex implement the refactor in services/billing/, then review and commit it.

3. Every dispatch runs the same review-first loop. The orchestrator writes a brief — self-contained, because the implementer has no orchestrator chat history — dispatches it with the bundled relay.mjs, and waits while the relay writes a structured result.json.
4. Review the diff and re-run the project's gates yourself; the README suggests pairing with the author's separate guard-skills repo.
5. Land it. You commit because, in the README's words, committing belongs to the reviewer — none of the relays ever commit.

Tips

• Read the table's write-access column before you trust a lane. Codex runs under --sandbox workspace-write, Claude Code under acceptEdits, Cursor under --force. Pi and Warp run full local tools with no sandbox or permission modes; for Command Code, --yolo is the only headless write state and there is no sandbox — the README's footnote says the brief's path list is guidance, not containment, a worktree isolates the checkout, and a container or another OS-enforced sandbox is required when writes outside the target tree are unacceptable.
• --read-only maps to each CLI's plan or read-only mode where one exists. Kimi and Warp have none; Grok's is best-effort, so the relay reports a readOnlyViolation tripwire instead of enforcing it.
• touchedFiles is a post-run git status. It cannot show ignored files, reverted edits, or writes outside the repository — it is what you review against, not a guarantee.
• Aider is the one implementer that commits by default; the relay always passes --no-auto-commits and --no-dirty-commits, and neither is configurable through it.
• Project lane config is content-bound to the setup approval: cloned or edited project lanes fail closed until re-approved, and the wrong implementer skill for a lane fails loud.
• Codex as orchestrator? Install codex-background separately; its bootstrap.mjs backs up and updates $CODEX_HOME/config.toml and registers the installed *-delegate skills behind a local MCP server.

When Not to Use This

This is a loop, not a forwarder, and it is not a subagent system: the README positions it as complementary to a vendor's own plugin or subagents, which coordinate inside one agent, while this keeps the contract portable across orchestrators with the commit on the reviewer. If what you want is coordination inside one agent, that is the job the README assigns to a vendor's own plugin or subagents. Also read the verification section as written: the full delegate → review → commit loop is designed for and run on Claude Code; other orchestrators are, in the README's words, designed-for but unproven, and opencode-delegate and vibe-delegate are contract-tested only, with no end-to-end run recorded.

See the leaderboard for more skills.