💡 Why CodexPotter
CodexPotter continuously reconciles code base toward your instructed state (Ralph Wiggum pattern):
- 🤖 Codex-first — Codex subscription is all you need; no extra LLM needed.
- 🧭 Auto-review / reconcile — Review and polish multi rounds until fully aligned with your instruction.
- 💦 Clean-room — Use clean context in each round, avoid context poisoning, maximize IQ.
- 🎯 Attention is all you need — Keep you focused on crafting tasks, instead of cleaning up unfinished work.
- 🚀 Never worse than Codex — Drive Codex, nothing more; no business prompts which may not suit you.
- 🧩 Seamless integration — AGENTS.md, skills & MCPs just work™ ; opt in to improve plan / review.
- 🧠 File system as memory — Store instructions in files to resist compaction and preserve all details.
- 🪶 Tiny footprint — Use <1k tokens, ensuring LLM context fully serves your business logic.
- 📚 Built-in knowledge base — Keep a local KB as index so Codex learns project fast in clean contexts.
👀 How does it work
𝒀𝑶𝑼𝑹 𝑷𝑹𝑶𝑴𝑷𝑻:
𝘚𝘪𝘮𝘱𝘭𝘪𝘧𝘺 𝘵𝘩𝘦 𝘲𝘶𝘦𝘳𝘺 𝘦𝘯𝘨𝘪𝘯𝘦 𝘣𝘺 𝘧𝘰𝘭𝘭𝘰𝘸𝘪𝘯𝘨 ...
│
│
codex: Work or review according to MAIN.md │
┌──────────────────────────┐ │
│ │ ▼
┌─────────┴─────────┐ ┌──────────▼────────┐ ┌───────────────────┐
│ CodexPotter │ │ codex │◄─────►│ MAIN.md │
└─────────▲─────────┘ └──────────┬────────┘ └───────────────────┘
│ │
│ Work finished │
└──────────────────────────┘
✨ What's New
- May 01, 2026: We now have a V2 version for Codex Desktop users! Checkout v2 branch (experimental) 🎉
⚡️ Getting started
1. Prerequisites: ensure you have codex CLI locally. CodexPotter drives your local codex to perform tasks.
2. Install CodexPotter via npm or bun:
# Install via npm
npm install -g codex-potter
# Install via bun
bun install -g codex-potter
3. Run: Start CodexPotter in your project directory, just like Codex:
# --yolo is recommended to be fully autonomous
codex-potter --yolo
⚠️ Note: Unlike Codex, every follow up prompt turns into a new task, not sharing previous contexts. Assign tasks to CodexPotter, instead of chat with it.
⚠️ Note: CodexPotter is not a replacement for codex, because CodexPotter is a loop executor — it executes tasks instead of chatting with you. See below for details.
Tips
Prompt Examples
✅ tasks with clear goals or scopes:
- "port upstream codex's /resume into this project, keep code aligned"
✅ persist results to review in later rounds:
- "create a design doc for ... in DESIGN.md"
❌ interactive tasks with human feedback loops:
CodexPotter is not suitable for such tasks, use codex instead:
- Front-end development with human UI feedback
- Question-answering
- Brainstorming sessions
Howto
Ask followup questions in codex
Just pass the project file to codex, like:
based on .codexpotter/projects/2026/03/18/1/MAIN.md,
please explain more about the root cause of the issue
Plan and execute
Simpliy queue two tasks in CodexPotter, one is plan, one is implement, CodexPotter will execute one by one, for example:
Task prompt 1 (CodexPotter):
Analyze the codebase, research and design a solution for introducing subscription system.
Output plan to docs/subscription_design.md.
Your solution should meet the following requirements: ...
Do not implement the plan, just design a good and simple solution.
↑ Your existing facility to write good plans will be utilized, including skills, plan doc principles in AGENTS.md, etc. Writing plan to a file is CRITICAL so that the plan can be iterated multiple rounds and task 2 can pick it up.
Task prompt 2 (CodexPotter):
Implement according to docs/subscription_design.md
Make sure all user journeys are properly covered by e2e tests and pass.
If you even don't know what you are designing for, just discuss with codex to carry out a basic plan first, then use CodexPotter to continously polish and implement it.
Configuration
Other Features
-
--xmodel(experimental): Use gpt-5.2 first, then use gpt-5.5 to cross review gpt-5.2's work in later rounds. In clear coding tasks this may produce better results than only using gpt-5.2 or gpt-5.5. -
/yolo: Toggle whether YOLO (no sandbox) is enabled by default for all sessions. -
/listorctrl+l: View all projects (tasks) and their results.
Roadmap
- Skill popup
- Resume (history replay + continue iterating)
- Better handling of stream disconnect / similar network issues
- Agent-call friendly (non-interactive exec and resume)
- Interoperability with codex CLI sessions (for follow-up prompts)
- Better plan / user selection support
- Better sandbox support
Development
# Formatting
cargo fmt
# Lints
cargo clippy
# Tests
cargo nextest run
# Build
cargo build
Community & License
- This project is community-driven fork of openai/codex repository, licensed under the same Apache-2.0 License.