Consigliere

Consigliere gives Claude Code a second brain. Codex GPT-5.6 Sol plans the work as a read-only advisor, and Claude Opus writes the code as the executor. Two models, each doing what it is strongest at.

The problem

When one model plans and builds in the same pass, it locks onto an approach fast and builds the wrong thing well. Splitting the roles fixes that: a planner that only reasons, a builder that only executes.

How it works

Your prompt goes to Sol first. It plans the approach, reads the code with tools like git diff and ripgrep, and hands back a plan. You approve it, Opus implements it, and Sol reviews the diff. Every finding is labeled adopt, discuss, style, or over-engineered, and all of them reach you.

What made it real

The idea was easy. The failure modes were the work.

Read-only, not on trust. The advisor runs in a sandbox with no write access, so it physically cannot edit your files. Planning stays planning.

A watchdog for hangs. Codex used to stall on a web search for 40 minutes. Now a background job is cancelled after five minutes of silence and control returns to Claude.

Honest reviews. Nothing is filtered out. Every finding is categorized and shown to you, and you decide what ships.

One command, any OS. A Node installer copies the hooks and merges into settings.json without touching your existing setup, backing up everything it edits.

Stack

It runs on your existing ChatGPT Plus login through the Codex plugin. No OpenAI API key, no proxy, no second subscription. MIT licensed.

Install with node install.mjs. Source is on GitHub. The idea was easy. The failure modes were the work.

Read-only, not on trust. The advisor runs in a sandbox with no write access, so it physically cannot edit your files. Planning stays planning.

A watchdog for hangs. Codex used to stall on a web search for 40 minutes. Now a background job is cancelled after five minutes of silence and control returns to Claude.

Honest reviews. Nothing is filtered out. Every finding