Phone a Friend

A Claude Code plugin that lets you "phone a friend" by calling OpenAI's Codex CLI for a second opinion. Supports bidirectional loops: Claude Code can call Codex, and Codex can call Claude Code back, with depth limiting to prevent infinite recursion.

Prerequisites

• Claude Code installed
• Codex CLI installed (npm install -g @openai/codex)
• OpenAI API key configured for Codex

Install

# From GitHub
claude plugins install --url https://github.com/orban/phone-a-friend

# From a local checkout
claude plugins install --path /path/to/phone-a-friend

# Or use --plugin-dir for session-only
claude --plugin-dir /path/to/phone-a-friend

Usage

Slash command

/phone-a-friend "What's the best way to handle concurrent writes to this database?"
/phone-a-friend "Review this error and suggest fixes" --mode critique
/phone-a-friend "Is this migration safe?" --mode verify
/phone-a-friend "What other approaches should I consider?" --mode explore
/phone-a-friend --input-file /tmp/diff.txt --mode critique

Automatic (skill-based)

Claude Code will automatically recognize when a second opinion would help and offer to phone Codex.

Loop example

Claude Code asks Codex, Codex asks Claude back:

You: /phone-a-friend "Is this SQL query optimal for 10M rows?"

Claude Code → Codex: "Analyze this SQL query for performance at 10M rows: SELECT ..."
Codex → Claude (if needed): "What indexes exist on this table?"
Claude → back to Codex: "The table has indexes on id, created_at, and user_id"
Codex → final answer back to Claude Code
Claude Code: synthesizes both perspectives for you

Configuration

Environment Variable	Default	Purpose
PHONE_A_FRIEND_DEPTH	0	Current loop depth (managed automatically)
PHONE_A_FRIEND_MAX_DEPTH	3	Max depth (inherited by child hops)
PHONE_A_FRIEND_CHAIN	(auto)	Call chain trace, e.g. claude>codex>claude

Flag	Default	Purpose
--mode	(general)	explore, critique, verify, or omit
--input-file	—	Attach file content (diffs, code, logs)
--model	(codex default)	Override Codex model
--max-depth	3	Maximum cc<->codex loop depth
--timeout	120	Seconds before killing a hung call
--working-dir	$(pwd)	Codex working directory

How the Loop Works

Depth 0: Claude Code (this session)
  │
  ├─ Runs: call-codex.sh "prompt"
  │
  Depth 1: Codex processes request
  │  │
  │  ├─ (optional) Runs: call-claude.sh "sub-question"
  │  │
  │  Depth 2: Claude Code answers sub-question
  │     │
  │     └─ (optional) Runs call-codex.sh again if depth < max
  │
  └─ Returns synthesized response

All hops go through the helper scripts, which enforce depth limits, timeouts, call-chain tracking, and proper error handling.

Reverse Direction: Codex Calling Claude

The scripts/call-claude.sh script is included for Codex to call Claude Code back. Codex receives instructions in its prompt about how to use it.

Files

phone-a-friend/
  commands/
    phone-a-friend.md     # /phone-a-friend slash command
  skills/
    phone-a-friend/
      SKILL.md            # Auto-detected skill
  scripts/
    call-codex.sh         # Claude Code → Codex
    call-claude.sh        # Codex → Claude Code (reverse)
  README.md