Route tasks to Claude, Codex, or Cursor—then review and commit

What it does

This skills package turns a shelf of incompatible AI coding CLIs—Claude Code, OpenAI Codex, Cursor, Aider, Grok, and nine others—into uniform subcontractors for an orchestrating agent. You dispatch a self-contained task to a specific implementer (or to a named lane like feature or tests), the relay runs it in a separate process, and you get back a diff and a structured report. You review, rerun the gates, and commit yourself; the relay is physically incapable of landing code.

The interesting bit

The entire trust model is built on deliberate boredom. Every relay is a single Node script using only built-ins—zero dependencies, zero network calls, zero telemetry—so the only thing it can leak is the command it forwards. By refusing to abstract away the implementer’s own flags and permission modes, it keeps the weirdness visible instead of pretending all agents behave alike.

Key highlights

• Supports 14 implementer CLIs, from mainstream tools like codex and claude to newer entrants like grok, kimi, and vibe
• Fleet lanes let you bind kinds of work to specific agents—ui to Cursor, tests to Codex—without hard-coding flags in every prompt
• Relays speak a common delegate-relay.result.v1 contract (status, exitCode, touchedFiles, session ID) so swapping agents doesn’t change your review loop
• Project-level fleet config is content-bound to explicit approval; clone or edit a lane and it fails closed until you re-approve
• The setup skill discovers what is already installed on your machine rather than pulling down new binaries

Caveats

• Several agents lack reliable read-only headless modes; Grok in particular cannot be prevented from writing, so the relay only reports a readOnlyViolation tripwire after the fact
• Cline does not support headless JSON resume through the relay, and some agents (Kimi, Pi, Warp) offer no CLI-enforced read-only mode at all
• Aider’s upstream defaults auto-commit and dirty-commit, which the relay has to forcibly disable on every run

Verdict

Worth installing if you already use multiple AI coding agents and want an orchestrator to route tasks without surrendering git. Skip it if you live inside one IDE with one agent, or if you expect the abstraction to sand off every implementer’s sharp edges.