codex 命令

常用示例

说明

Codex is OpenAI's official terminal-based coding agent. It launches into a full-screen terminal UI that can read your repository, suggest and implement code changes, and execute commands while you iterate conversationally. The tool maintains security through OS-level sandboxing with three safety modes: read-only (consultative, no changes), workspace-write (default, asks before touching files outside workspace), and danger-full-access (unrestricted, use with caution). Model-generated commands run in isolated environments using macOS seatbelt or Linux Landlock. Codex integrates with ChatGPT Plus, Pro, Business, and Enterprise subscriptions, or can be used with API credits. It supports cloud tasks for offloading complex work and provides MCP integration for extended capabilities.

参数

--model, -m _MODEL_

Override the default model (e.g., gpt-5.2-codex, codex-mini-latest)

--sandbox, -s _POLICY_

Sandbox policy: read-only, workspace-write, or danger-full-access

--ask-for-approval, -a _MODE_

Control approval timing: untrusted, on-failure, on-request, or never

--full-auto

Enable low-friction mode with on-request approvals and workspace-write sandbox

--yolo

Disable approvals and sandboxing (use only in hardened environments)

--cd, -C _PATH_

Set working directory before processing

--add-dir _PATH_

Grant additional directories write access

--image, -i _PATH_

Attach image files to the initial prompt

--oss

Use local open-source model via Ollama

--search

Enable web search capability

--profile, -p _NAME_

Load configuration profile from config.toml

--help

Display help information