Codez

A Codex-compatible runtime fork for local agents, token-aware context control, App Server v2, and plugin hooks.

Codez keeps upstream Codex compatibility where it matters, then layers focused runtime experiments on top.

Compatibility · RTK Plugin · Getting Started · Config

Codez is a fork of OpenAI Codex. The goal is not to hide that lineage: upstream Codex remains the base CLI, protocol, package shape, and user-facing mental model. Codez is the fork layer for local-agent runtime work that benefits from moving faster than upstream.

Why Codez Exists

• keep a Codex-compatible CLI/runtime while experimenting with local-agent needs
• reduce wasted tokens from stale tool, reasoning, and context history
• support plugin and hook workflows that can harden shell usage
• expose App Server v2 surfaces for local clients and future gateways
• make context-pressure and prompt-history behavior easier to evolve
• provide a clear core layer for future gateway projects without baking a specific gateway into the runtime

Core Additions

Codez is not only a renamed README around upstream Codex. The fork carries runtime work aimed at long local-agent sessions:

• prompt-history pruning before model sampling: stale context messages, older reasoning items, tool calls, and tool outputs can be removed before the next request while preserving the active turn and image-output dependencies
• goal-continuation pruning: automatic goal follow-up turns are treated as fresh prompt boundaries, so previous tool-heavy work can be trimmed while the active goal objective remains visible
• compaction-aware pruning: the remote/autocompact path can apply the same pruning and trim function-call history before compaction, reducing context pressure before a summarization turn
• App Server v2: a local app/server protocol surface for richer clients, gateway experiments, thread operations, hook/catalog inspection, and command event flows
• plugin hook compatibility: Codez keeps plugin-loaded hook paths usable for PreToolUse workflows, so optional safety plugins such as RTK can run as normal runtime extensions

What Is Different

Codez keeps upstream names such as codex, @openai/codex, and Codex protocol terms where compatibility matters. Fork-specific docs use the Codez name when describing additions, release projections, or stack positioning.

The public stack shape is:

Codez runtime
  -> optional RTK Codex Plugin for shell token-safety
  -> Telegram gateway layer later (Teledex coming next)

RTK is already published separately:

https://github.com/Krablante/rtk-codex-plugin

The Telegram gateway layer is not linked here yet because its public repo has not been published.

Quick Start

Packaged Codez releases are a publication decision. Until then, use the source tree directly:

git clone https://github.com/Krablante/codez
cd codez
pnpm install
just install
just codex --help

For upstream Codex usage, package names and official setup are still documented by OpenAI:

• Codex documentation
• Upstream repository

Read Next

• Compatibility matrix
• Optional RTK plugin integration
• Installing and building
• Configuration
• Contributing

License and Attribution

Codez is distributed under the same Apache-2.0 license as upstream Codex. OpenAI Codex remains the upstream project; Codez is an independent fork projection and is not presented as an official OpenAI release.