This paper describes the "Codex plugin for Claude Code," a software integration designed to bridge OpenAI's Codex code intelligence capabilities directly into the Claude Code development environment. The plugin's primary objective is to allow Claude Code users to leverage Codex functionalities for tasks such as code review, bug investigation, and general code delegation without leaving their existing workflow.

The core methodology revolves around a command-line interface (CLI) driven interaction within Claude Code, utilizing a set of /codex: slash commands. The plugin operates by wrapping and delegating tasks to the user's local Codex CLI and app server installation, ensuring consistency with existing Codex configurations, authentication, and local environment.

Key functionalities provided include:

1. Code Review (/codex:review): This command initiates a standard, read-only code review. It can be applied to current uncommitted changes or used to compare a branch against a specified base branch (e.g., --base main). It supports asynchronous execution via --background and synchronous waiting via --wait. This review is not steerable and does not accept custom focus text.

2. Adversarial Code Review (/codex:adversarial-review): This advanced review mode provides a steerable mechanism to critically question implementation choices, design decisions, and underlying assumptions. It helps pressure-test for potential tradeoffs, failure modes, and explore alternative approaches. Like /codex:review, it supports --base <ref>, --background, and --wait, but uniquely accepts additional focus text to guide the adversarial questioning (e.g., "challenge whether this was the right caching and retry design"). This command is also read-only and does not modify code.

3. Task Delegation (/codex:rescue): This command delegates a specified task to Codex via the codex:codex-rescue subagent. Tasks can include bug investigation, fixing issues, or continuing previous Codex work. It supports execution in the background (--background), synchronous waiting (--wait), resuming a previous task (--resume), or starting fresh (--fresh). Users can specify the Codex model (--model e.g., gpt-5.4-mini, spark which maps to gpt-5.3-codex-spark) and the reasoning effort (--effort e.g., medium). If model and effort are omitted, Codex defaults are used.

4. Job Management (/codex:status, /codex:result, /codex:cancel):
   • /codex:status: Provides an overview of active and recently completed Codex jobs within the current repository.
   • /codex:result: Displays the final output of a completed Codex job, including the Codex session ID, which allows direct resumption of the work within the standalone Codex application via $codex resume <session-id>$.
   • /codex:cancel: Terminates an ongoing background Codex job.

5. Setup and Configuration (/codex:setup): This command verifies the installation and authentication status of Codex. It can offer to install Codex if detected as missing and npm is available. A significant feature managed by /codex:setup is the "review gate." When enabled via --enable-review-gate, the plugin intercepts Claude's response using a "Stop hook," triggering a targeted Codex review. If this review identifies issues, Claude's output is blocked, requiring it to address the detected problems, thus creating a feedback loop. This mechanism, while powerful for ensuring quality, is cautioned to potentially lead to long loops and high resource consumption.

The plugin's technical operation relies on the fact that it does not introduce a separate Codex runtime but rather acts as an orchestrator for the existing local Codex installation. This design choice ensures that all existing Codex configurations (e.g., in ~/.codex/config.toml or project-level .codex/config.toml files, specifying model and model_reasoning_effort), authentication states (using ChatGPT subscription or OpenAI API key), and local environment variables are honored and leveraged seamlessly by the plugin. The integration ensures that any work delegated via the plugin can be directly resumed and further managed within the standalone Codex application.