Codex CLI 八大隐藏神技：终端党必看的信息差

flowchart TB
    subgraph init["🚀 启动阶段"]
        A["codex 启动"] --> B["读取 AGENTS.md<br/>(项目级指令)"]
        B --> C["加载 config.toml<br/>(用户/项目配置)"]
        C --> D["选择 Profile<br/>(fast/thorough)"]
    end

    subgraph modes["⚡ 四种执行模式"]
        E["/plan<br/>只思考不执行"]
        F["Default Mode<br/>执行前确认"]
        G["--full-auto<br/>全自动执行"]
        H["/goal 🔥<br/>多小时代理循环"]
    end

    subgraph skills["🧩 技能扩展层"]
        I["内置 Skills<br/>/browser-use<br/>/image-gen<br/>/pdf"]
        J["自定义 Skills<br/>Markdown 编写"]
        K["MCP Servers<br/>外部工具连接"]
    end

    subgraph output["📦 输出与集成"]
        L["codex exec<br/>非交互/CI/CD"]
        M["codex review<br/>代码审查"]
        N["codex apply<br/>应用 diff"]
    end

    D --> E
    D --> F
    D --> G
    D --> H
    F --> I
    F --> J
    F --> K
    G --> L
    G --> M
    G --> N

    style H fill:#ff6b6b,color:#fff
    style G fill:#ffd93d,color:#333
    style L fill:#4d96ff,color:#fff

# 1. 编辑 config.toml 启用 /goal（默认关闭）
vim ~/.codex/config.toml

# 添加以下内容：
[experimental]
goal_mode = true

# 2. 在 Codex 中启动
codex
> /goal build a complete extraction shooter game with auto-generated assets

# 方法一：用内置命令自动生成
codex
> /init
# Codex 会扫描项目结构，自动生成 AGENTS.md

# 方法二：手动创建
cat > AGENTS.md << 'EOF'
# My Project

## Build Commands
- Test: `npm test`
- Build: `npm run build`
- Lint: `npm run lint -- --fix`

## Rules
- Use TypeScript strict mode for all new files
- Write unit tests for every new function
- Never commit directly to main branch
- All API endpoints must have OpenAPI docs

## Architecture
- Frontend: React + Tailwind CSS
- Backend: Express + PostgreSQL
- Auth: JWT with refresh tokens
EOF

> Based on our discussion, generate an AGENTS.md for this project

# ~/.codex/config.toml

# 快速模式：简单任务用
[profiles.fast]
model = "gpt-5.4-mini"
model_reasoning_effort = "low"
service_tier = "fast"

# 深度模式：架构级工作用
[profiles.thorough]
model = "gpt-5.4"
model_reasoning_effort = "xhigh"

# 代码审查模式
[profiles.review]
model = "gpt-5.4"
model_reasoning_effort = "high"
sandbox_mode = "read-only"

# 快速修个 typo
codex -p fast "fix the typo in README"

# 深度架构重设计
codex -p thorough "redesign the auth module"

# 只读代码审查
codex -p review --uncommitted

# 1. 创建技能目录
mkdir -p ~/.codex/skills

# 2. 创建自定义技能
cat > ~/.codex/skills/code-review.md << 'EOF'
# Code Review Skill

## Instructions
1. Run `git diff` to see all changes
2. Check for: security issues, performance problems, missing tests
3. Check code style consistency with project conventions
4. Generate a review summary with severity levels (critical/warning/info)
5. Suggest specific fixes for each issue found

## Output Format
- Start with overall assessment
- List issues by severity
- Include file:line references
- End with actionable next steps
EOF

# 3. 在对话中调用
codex
> $code-review

flowchart LR
    A["~/.codex/skills/<br/>个人级<br/>所有项目可用"] --> D[调用方式<br/>$skill-name]
    B[".agents/skills/<br/>项目级<br/>仅当前项目"] --> D
    C["/etc/codex/skills/<br/>系统级<br/>所有用户"] --> D

    style A fill:#6bcb77,color:#fff
    style B fill:#4d96ff,color:#fff
    style C fill:#ffd93d,color:#333

> I just did a complex database migration. Create a skill from what we just did.

# 创建规则目录
mkdir -p ~/.codex/rules

# 创建默认规则
cat > ~/.codex/rules/default.rules << 'EOF'
# 禁止直接 push
prefix_rule(
    pattern = ["git", "push"],
    decision = "forbidden",
    justification = "No direct pushes - use PR workflow"
)

# 危险操作需确认
prefix_rule(
    pattern = ["rm", "-rf"],
    decision = "prompt",
    justification = "Destructive operation needs confirmation"
)

# 测试命令自动批准
prefix_rule(
    pattern = ["npm", "test"],
    decision = "allowed",
    justification = "Tests are safe to run"
)

# Docker 命令需确认
prefix_rule(
    pattern = ["docker"],
    decision = "prompt",
    justification = "Docker commands affect system state"
)
EOF

# 测试规则（不会真的执行）
codex exec policy check --pretty -- git push origin main
# 输出: FORBIDDEN - No direct pushes - use PR workflow

# 基本用法：非交互执行任务
codex exec "fix the failing tests"

# 输出 JSON（适合脚本解析）
codex exec --json "list all API endpoints"

# 保存结果到文件
codex exec -o result.txt "summarize the architecture"

# 强制 JSON Schema 输出
codex exec --output-schema schema.json "extract metadata"

# 恢复上一次对话继续执行
codex exec resume --last "now add error handling"

# CI/CD 全自动模式
codex exec --full-auto "run tests and fix all failures"

# .github/workflows/auto-fix.yml
name: Auto Fix Failures
on:
  workflow_run:
    workflows: ["Tests"]
    types: [completed]
    result: failure

jobs:
  auto-fix:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Auto-fix with Codex
        env:
          CODEX_API_KEY: ${{ secrets.CODEX_API_KEY }}
        run: |
          codex exec --full-auto \
            --config model=gpt-5.4-mini \
            "run tests and fix all failures"
      - name: Create PR
        run: |
          git checkout -b auto-fix-$(date +%s)
          git add -A
          git commit -m "auto-fix: test failures"
          gh pr create --title "Auto-fix: test failures" --body "Automated fix by Codex"

# 1. 安装 Ollama（如果没装的话）
curl -fsSL https://ollama.ai/install.sh | sh

# 2. 拉一个编码模型
ollama pull qwen3-coder:30b

# 3. 用 --oss 标志启动 Codex
codex --oss --local-provider ollama

# 或者用 LM Studio
codex --oss --local-provider lmstudio

codex
> /browser-use
> Test the dashboard at localhost:3000, find all UI bugs and fix them

# my-build-skill.md
## Instructions
1. Build the feature as described
2. Run /browser-use to test the UI
3. Fix all bugs found during browser testing
4. Only present the result after browser test passes

> /personality
# 选择 Pragmatic（务实）模式
# 效果：Codex 不再说"好问题！"，直接给答案和下一步

> /compact
# 对话太长时，压缩历史释放 token 空间
# 类似 Claude Code 的 /compact

# 分析截图
codex -i screenshot.png "what's wrong with this UI?"

# 对比两张图
codex -i before.png,after.png "compare these two designs"