Codex Factory plugin
Codex Factory runs a complex Codex project as a controlled process: plans, roles, source-bound memory and independent verification
- 90 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Installs project-level hooks and runs several Codex subagents that write files and execute commands in the project
- Stores project memory and a GLM search key in local files, and with search on, queries go to an external service
Install
- Works in
Manual install
git clone https://github.com/yueheng-rgb/codex-factory.git C:\Codex_App_FactoryThen `packages/factory-cli/install.ps1` with `-ProjectRoot`. Node.js 24 and PowerShell are required.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Installs project-level hooks and runs several Codex subagents that write files and execute commands in the project
- Stores project memory and a GLM search key in local files, and with search on, queries go to an external service
This is third-party code. Review the repository files before installing.
What it does
A local control plane for developers who build non-trivial software in Codex or compatible runtimes. It turns a long fragile chat into a process: task routing, trusted context, native subagents and verification by independent evidence. By the authors' rule a PASS without an artifact does not count. Evidence means real files, command exit codes, hashes and receipts. V5 installs on Windows with `packages/factory-cli/install.ps1` and the `-MultiAgent`, `-Search` and `-MaxThreads` flags, and `factoryctl doctor` checks the state. Multi-agent mode is opt-in: the main Codex agent splits the work itself, starts resident roles and temporary agents and runs verification waves. Memory selects only current entries by task and role, bound to a source and a hash. GLM search is optional, and its key comes from `ZHIPUAI_API_KEY` or an ignored `.codex-factory/secrets.env`. Project-level hooks check the subagent lifecycle.
Who it is for. Codex developers whose large projects lose the thread in long conversations and who need verifiable evidence that tasks are done.
Good fit when
- A multi-week project needs its context stored with source binding and not in chat history
- An agent reports done and completion must be confirmed by files, hashes and command exit codes
- You want several subagents with role assignment and verification waves
- You want an offline walkthrough of scenarios with `npm run demo:tour` and no model calls
Not a fit when
- You need a simple edit in one file: the control layer is overkill
- The environment has no Windows and PowerShell: the V5 installer is written for them
- You need OS-level protection: `scope_guard` detects out-of-bounds writes but does not replace access rights or a separate git tree
Example request
Set up Codex Factory for the project in C:\Projects\my-app with multi-agent mode on four threads and check factoryctl doctorLimitations
- V5 is described as a preview. The installer targets Windows and PowerShell and needs Node.js 24 or newer and a Codex runtime that supports project-level customization.
- `scope_guard` finds missed writes and writes outside a wave but gives no per-agent author and does not replace OS permissions.
- Project hooks must be reviewed and trusted in `/hooks` on first run or after a change, under Codex security rules.
- The repository is large and holds many historical V4 reports, and its last commit is dated September 2026.
How to disable. Uninstall from `packages/factory-cli` following `docs/CONTROL_PLANE_V5_GUIDE.zh-CN.md`, or delete the `.codex-factory` folders and the added hooks from the project.
FAQ
How do I install V5?
On Windows clone the repository and run `packages/factory-cli/install.ps1` with `-ProjectRoot`, `-MultiAgent`, `-Search none` and `-MaxThreads 4`. Afterwards `factoryctl doctor --project <path> --json` checks readiness.
What is the rule about an artifact and a PASS?
A result counts only with independent evidence: real files, a command exit code, a hash or a receipt. An agent's claim of completion is not evidence.
Is multi-agent mode required?
No, it is opt-in. Without it Factory works with plans, role skills, memory and verification on a single agent.
Where does the GLM search key come from?
From the `ZHIPUAI_API_KEY` variable or the ignored `.codex-factory/secrets.env` file in the project. Search is switched on with the separate `-Search` flag.
How can I try it without a model?
In `packages/factory-cli` run `npm run demo:tour`: it shows three scenarios offline, outcome comparison, reuse of corrections and hypothesis-driven debugging. It is a tour and not a real project delivery.
Discussion
No questions or reviews yet.
Related
Pluginsoh-my-codex
oh-my-codex (OMX)
A layer on top of Codex CLI: one flow from clarifying a task to verification, agent teams, hooks, a HUD and state in .omx/
Codex on Crack splits Codex work across models: the lead plans, a worker implements and the lead reviews the result.
OpenAI's coding agent that runs locally in your terminal, with IDE and desktop versions
A Codex plugin that builds a team of subagents and skills for a project: an orchestrator splits tasks and checks the result.
CodexPotter loops Codex with a clean context each round until the code matches the goal, as a replacement for the built-in /goal.
Codex Astra + Luna Orchestrator
Codex Astra Orchestrator + Luna Subagents
Codex profiles where GPT-6 Astra orchestrates and reviews while GPT-5.6 Luna runs the explorer, worker, tester and researcher subagents
README in short
The README presents Codex Factory as an engineering reliability layer for coding agents. It has a short pitch, the preview V5 with installation and boundaries, a quick start, historical V4 verification, architecture and a limitations section. The main guides are written in Chinese.