claudex

Claudex

A Claude Code plugin for multi-AI deliberation between Claude and Codex.

Claude reasons inline, a persistent Codex partner runs in the background, and you watch both sides of the debate stream live in your terminal — like overhearing a technical argument between two senior engineers.

Install

/plugin install bhaskar-melkani/claude-codex

Registers automatically:

• /deliberate and /deliberate-template skills
• /council skill for parallel multi-track analysis
• Codex MCP server wiring
• Auto-deliberation hook on /plan

Prerequisites

• Claude Code installed
• Codex CLI installed as codex and authenticated

---

Usage

Direct deliberation

/deliberate "Should we use Prisma or Drizzle for the new service?"

From a template (guided topic builder)

/deliberate-template

Picks from 12 common decision types (REST vs GraphQL, SQL vs NoSQL, Monolith vs Microservices, etc.) and asks 2-3 questions to build a focused topic before launching.

Multi-track council (parallel analysis)

/council "Should we migrate to microservices?"

Decomposes the decision into 2-4 orthogonal tracks (cost, deployment, team structure, performance) and runs each as a simultaneous Claude+Codex deliberation. Results are synthesized into a unified recommendation.

Automatic on /plan

/plan should we switch from Postgres to CockroachDB?

The hook detects /plan and injects the full deliberation workflow automatically.

Automatic on plan mode (shift+tab)

Press shift+tab in Claude Code to enter plan mode. The hook fires immediately and injects a fast 2-round deliberation before Claude produces the plan — no extra command needed. Claude deliberates with Codex, then leads with the plan. The deliberation summary is shown in 2–3 bullets; the full transcript is hidden unless you ask.

---

What the terminal looks like

================================================================
  DELIBERATION STARTING
================================================================
  Topic     : Should we use Prisma or Drizzle for the new service?
  Claude    : Opus 4.6  (inline)
  Codex     : GPT-5.4-mini  (background partner)
  Max rounds: 3  (min 2)
  Log       : ~/.claude/deliberations/2026-04-03-prisma-vs-drizzle-a4f1.md
================================================================

● Codex deliberation partner
  L Running in the background

================================================================
  ROUND 1 / 5
================================================================

  ┌─ Claude ──────────────────────────────────────────────────
  │ **Thesis:** Prisma is the right choice for this team because
  │ its type-safe client eliminates an entire class of runtime errors.
  │
  │ ### Position
  │ At ~50k writes/day with a TypeScript-first team and no prior ORM
  │ experience, Prisma's generated client is a significant productivity
  │ win...
  └───────────────────────────────────────────────────────────

  ┌─ Codex  ──────────────────────────────────────────────────
  │  (thinking...)
  ...
  │ **Thesis:** Drizzle's zero-overhead SQL approach is safer at scale.
  │
  │ ### Position
  │ Prisma's query engine adds ~20ms cold-start latency in serverless
  │ environments and its migration system has known pain points...
  └───────────────────────────────────────────────────────────

  Round 1 complete.
  Still disputed: migration DX, serverless cold-start overhead
----------------------------------------------------------------

---

Architecture

Main session (Claude, inline)           Codex partner (background agent)
─────────────────────────────────       ──────────────────────────────────
Phase 1: setup, spawn partner       ──► partner starts, awaits Round 1

Round 1:
  Claude position (streams live)
  SendMessage("codex-partner")      ──► calls mcp__codex__codex (GPT-5.4-mini)
                                        stores threadId internally
                        ◄── response ──
  Print Codex response (live)

Round 2–5:
  Claude responds to Codex
  SendMessage("codex-partner")      ──► calls mcp__codex__codex-reply
                        ◄── response ──   (reuses threadId from Round 1)
  Print Codex response (live)

Phase 3: tie-breaking, final plan
  SendMessage("DELIBERATION COMPLETE") ► partner exits

Key properties:

• One persistent agent for the whole deliberation (not a new agent per round)
• threadId lives inside the partner — no JSON extraction or passing between agents
• Both sides print live — you see the conversation as it happens
• Native Claude Code tree UI — partner shows as ● Running in the background

---

Skills

/deliberate

Core deliberation skill. Runs 2–5 structured rounds between Claude (Opus 4.6, inline) and Codex (GPT-5.4-mini, background partner).