Confer

Consult the other AI coding harness from inside your current session for a genuine second opinion. The peer sees none of this conversation — the value of this skill is packaging a clean cold-start briefing, invoking the peer's own CLI correctly, and surfacing its reply verbatim.

The peer runs as a fresh subprocess with its own session, auth, and config, so this works even nested (this skill invoked inside a Codex session that Claude itself spawned, or vice versa). Proceed normally.

This skill only ever shells out to an already-installed, already-authenticated peer CLI. It uses no SDK and carries no API keys of its own — it reuses whatever the peer harness is logged into. Shelling out to claude -p / codex exec is each vendor's officially-supported headless entry point.

Step 1 — Pick the direction and load the reference

You run inside one harness; you confer with the other. Detect which you are, then read the matching reference for the exact, current invocation — always open the reference before invoking; CLI flags and model names drift.

You usually already know your own identity — use it; the env var is the backstop: