Rubber Duck for Claude Code

>o)
(_>

Cross-model second opinion companion for AI coding agents.

Inspired by GitHub Copilot's Rubber Duck — the same idea, built for Claude Code using Codex CLI as the cross-model reviewer.

Like the classic debugging practice — you explain what you did, the duck listens, and sometimes it spots what you missed.

What it does

• Reviews the decisions your AI agent made — not code style
• Gives a plain-language take, like a colleague's quick read
• Ends with an actionable bottom line (ship it / fix X / rethink Y)
• Activates automatically at high-signal checkpoints or on demand

Example

>o)
(_>

Took a look at what you just did.

• The retry logic in fetcher.ts:82 doesn't back off — it'll hammer the endpoint on transient failures. This project uses p-retry elsewhere, probably worth using it here too.

• user.preferences can be null but you're accessing .theme directly on line 112. Added optional chaining.

Ship it after switching to p-retry.

Install

Prerequisites

Codex CLI must be installed and authenticated:

npm install -g @openai/codex
codex login

Via APM (recommended — automatic + on demand)

apm install chkp-roniz/cc-rubber-duck
apm compile --target claude --local-only

APM compiles the trigger criteria into your project's CLAUDE.md, so Claude sees them at session start and invokes the duck automatically at high-signal checkpoints.

Via Claude Code plugin (on demand only)

claude /plugin marketplace add chkp-roniz/cc-rubber-duck

Claude sees the skill description (which includes trigger hints), but automatic activation depends on the agent reading the description and deciding to invoke. Use /rubber-duck to invoke explicitly.

Manual

Copy skills/, commands/, and instructions/ into your project's .claude/:

cp -r skills/ .claude/skills/
cp -r commands/ .claude/commands/
cp -r instructions/ .claude/instructions/

Usage

On demand

Type /rubber-duck in Claude Code, or just say "rubber duck this", "second opinion", "what am I missing", etc.

Automatic (APM or manual install)

The duck activates on its own when Claude:

• Finishes a plan touching 3+ files
• Writes 100+ lines of meaningful changes
• Completes a test suite
• Gets stuck retrying the same fix
• Is about to do something destructive

How it works

1. Shows the duck logo
2. Gathers context (git diff, recent files, or conversation context)
3. Pipes it to codex exec with a decision-review prompt
4. Evaluates concerns against full codebase context
5. Presents a short, honest summary with an actionable bottom line

The duck runs in --sandbox read-only mode — it can't modify your code.

Project structure

rubber-duck/
├── plugin.json                    # Plugin manifest
├── instructions/
│   └── rubber-duck.instructions.md  # Trigger criteria (compiled into CLAUDE.md by APM)
├── skills/
│   └── rubber-duck/
│       └── SKILL.md               # Full skill definition (behavior + presentation)
├── commands/
│   └── rubber-duck.md             # /rubber-duck slash command
└── README.md

License

MIT