amElnagdy/delegate-skills

delegate‑skills – Orchestrating AI‑powered Code Implementers

What it is – A collection of skills (small Node‑JS scripts) that let you plug a variety of LLM‑driven code‑generation CLIs (Codex, Claude, Cursor, Aider, etc.) into a single, consistent workflow. The package does not contain an LLM itself; it discovers the CLIs you already have installed, groups them into named lanes (e.g., feature, tests, ui), and then dispatches a task to the appropriate CLI while keeping you in the review loop.

Why it matters – Developers often juggle several AI coding assistants, each with its own command‑line interface and quirks. delegate‑skills gives you a unified front‑end:

• One command (delegate‑setup) to auto‑detect the CLIs on your machine and generate a fleet configuration.
• Per‑lane delegation (*-delegate --lane <name>) that guarantees the same JSON result format, read‑only safety modes, and a mandatory manual review before any commit.
• No hidden network calls or telemetry – the scripts are pure Node built‑ins that only launch the chosen CLI and git.

How it works

1. Setup – Run $delegate-setup. It scans your PATH for supported CLIs, proposes a fleet (a map of lane → implementer + optional flags), shows you the full config, and writes it only after you approve.
2. Dispatch – Use a skill like $codex-delegate --lane feature or $cursor-delegate directly. The skill calls a tiny relay.mjs script that:
   • Writes a brief (the task description you provide).
   • Launches the target CLI in headless mode.
   • Captures its output, the git diff, and a result.json that follows the delegate‑relay.result.v1 contract (status, exitCode, touchedFiles, session id, etc.).
3. Review – You inspect the diff, run any project‑specific gates, and decide whether to commit. The relay never commits; committing is always a manual step.
4. Land – After approval you git commit the changes.

Supported implementers – The repo ships a ready‑made skill for each of the following CLIs (the table in the README lists their flags, read‑only modes, and resume capabilities):

• Aider, Antigravity (agy), 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, Z.AI ZCode.
• Each skill’s SKILL.md documents required installation, authentication, and any caveats (e.g., whether the CLI supports a true read‑only mode).

Safety guarantees

• No automatic commits – The relay only writes a diff; you must manually commit.
• Read‑only / plan modes – Most skills expose a flag that runs the CLI in a “dry‑run” where no filesystem‑changing tools are allowed.
• Permission transparency – The relay reports readOnlyViolation and readOnly flags so you know if a CLI tried to write when it shouldn’t.
• Zero runtime dependencies – Only Node built‑ins are used; the scripts do not contact any external service themselves.

Installation – Via the Vercel Skills CLI:

npx skills add amElnagdy/delegate-skills          # whole package
npx skills add amElnagdy/delegate-skills --skill delegate-setup   # just the setup skill
npx skills add amElnagdy/delegate-skills --skill codex-delegate   # a single implementer

You can install globally or per‑agent, and pin to a specific version tag.

Typical use‑case

# 1️⃣ Discover CLIs and create a fleet
$delegate-setup
# → prompts you, writes .skills/delegate-fleet.json

# 2️⃣ Ask an LLM to refactor a module, using the lane you assigned to Codex
$codex-delegate --lane feature "Refactor services/billing to use the new pricing model"

# 3️⃣ Review the diff that appears, run tests, then commit
git diff && npm test && git commit -am "refactor billing"

Who should use it – Developers who already rely on one or more AI coding assistants and want a reproducible, auditable way to route tasks to the best‑suited tool while keeping the final commit under human control.

---

All details above are taken directly from the repository’s README; no additional features have been inferred.