ai-coding-agents（AI 编码剂）

AI Coding Agents Skill

Expert knowledge for Codex CLI and Claude Code CLI — the two leading AI coding agents.

Note: This skill documents both tools for reference. VMark development primarily uses Claude Code CLI. The Codex CLI sections are retained for completeness and cross-tool workflows.

When to Use

• Orchestrating complex coding tasks via CLI
• Configuring MCP servers for either tool
• Setting up automation pipelines (CI/CD)
• Troubleshooting authentication or sandbox issues
• Comparing capabilities between agents
• Custom agent/subagent configuration

Quick Reference

Starting Sessions

Safety Modes

Model Selection

Codex CLI (OpenAI)

Installation

npm i -g @openai/codex
# or
brew install --cask codex

Authentication

codex login              # OAuth via ChatGPT
codex login --with-api-key   # Read API key from stdin
codex login status       # Check auth status
codex logout             # Remove credentials

Core Commands

codex - Interactive Mode

codex                           # Start TUI
codex "fix all TypeScript errors"  # With initial prompt
codex -i screenshot.png "explain"  # With image
codex --full-auto "refactor"    # Low-friction mode
codex --search "find docs"      # Enable web search

codex exec - Non-Interactive

codex exec "write tests"        # Run and exit
codex e "task"                  # Short alias
echo "task" | codex exec -      # From stdin
codex exec --json "task"        # JSONL output
codex exec -o result.txt "task" # Save to file
codex exec --output-schema schema.json "task"  # Validate output

codex resume - Continue Sessions

codex resume                    # Interactive picker
codex resume --last             # Most recent
codex resume --all              # Show all (any directory)
codex resume <session-id>       # Specific session
codex resume <id> "continue with this"  # With prompt

codex review - Code Review

codex review                    # Review current branch vs main
codex review --uncommitted      # Review uncommitted changes
codex review --base develop     # Against specific branch
codex review --commit abc123    # Review specific commit
codex review "focus on security"  # Custom instructions

codex apply - Apply Cloud Task

codex apply <task-id>           # Apply diff from cloud task

codex cloud - Cloud Tasks (Experimental)

codex cloud                     # Browse cloud tasks
codex cloud exec "task" --env <env-id>  # Submit task
codex cloud status <task-id>    # Check status
codex cloud diff <task-id>      # Show diff
codex cloud apply <task-id>     # Apply changes

codex mcp - MCP Server Management

codex mcp list                  # List servers
codex mcp list --json           # JSON output
codex mcp get <name>            # Server details
codex mcp add <name> -- npx my-server   # Add stdio server
codex mcp add <name> --url https://... # Add HTTP server
codex mcp add <name> --env API_KEY=xxx -- cmd  # With env vars
codex mcp remove <name>         # Remove server
codex mcp login <name> --scopes read,write  # OAuth for HTTP
codex mcp logout <name>         # Remove OAuth

codex sandbox - Run Sandboxed Commands

# macOS
codex sandbox macos -- npm test
codex sandbox seatbelt --full-auto -- ./script.sh

# Linux
codex sandbox linux -- npm test
codex sandbox landlock -- ./script.sh

codex completion - Shell Completions

codex completion bash >> ~/.bashrc
codex completion zsh >> ~/.zshrc
codex completion fish > ~/.config/fish/completions/codex.fish

Slash Commands (Interactive)

Configuration (~/.codex/config.toml)

model = "gpt-5-codex"
approval_policy = "on-request"

[sandbox]
mode = "workspace-write"

[features]
web_search = true

[profiles.ci]
model = "gpt-4.1"
approval_policy = "never"

Global Flags

-m, --model <MODEL>          Model selection
-s, --sandbox <MODE>         read-only|workspace-write|danger-full-access
-a, --ask-for-approval <P>   untrusted|on-failure|on-request|never
-c, --config <KEY=VALUE>     Override config
-C, --cd <DIR>               Working directory
-i, --image <FILE>           Attach image(s)
-p, --profile <NAME>         Config profile
--full-auto                  Low-friction mode
--yolo                       Bypass all safety (DANGEROUS)
--search                     Enable web search
--add-dir <DIR>              Grant additional write access
--enable <FEATURE>           Enable feature flag
--disable <FEATURE>          Disable feature flag
--oss                        Use local OSS model

Claude Code CLI (Anthropic)

Installation

npm install -g @anthropic-ai/claude-code

Authentication

claude                      # First run prompts login
claude setup-token          # Set up long-lived token
# Requires Claude Pro/Max subscription OR API key

Core Commands

claude - Interactive Mode

claude                          # Start REPL
claude "explain this project"   # With prompt
claude -c                       # Continue last conversation
claude -r "session-name"        # Resume by name/ID
claude --model opus             # Select model
claude --chrome                 # Enable Chrome integration
claude --ide                    # Auto-connect to IDE

claude -p - Print Mode (Non-Interactive)

claude -p "explain this function"   # Query and exit
cat file | claude -p "explain"      # Process piped input
claude -p --output-format json "q"  # JSON output
claude -p --output-format stream-json "q"  # Streaming JSON
claude -p --max-turns 3 "task"      # Limit agent turns
claude -p --max-budget-usd 5 "task" # Spending limit
claude -p --json-schema '{...}' "q" # Validate output schema

claude mcp - MCP Server Management

claude mcp list                 # List servers
claude mcp get <name>           # Server details
claude mcp add <name> <cmd>     # Add stdio server
claude mcp add -t http <name> <url>  # Add HTTP server
claude mcp add -e KEY=val <name> -- cmd  # With env vars
claude mcp add -H "Auth: Bearer x" <name> <url>  # With headers
claude mcp add -s project <name> <cmd>  # Project scope
claude mcp remove <name>        # Remove server
claude mcp serve                # Run as MCP server
claude mcp add-from-claude-desktop   # Import from desktop app
claude mcp reset-project-choices     # Reset approvals

claude plugin - Plugin Management

claude plugin list              # List plugins
claude plugin install <name>    # Install plugin
claude plugin install <name>@marketplace  # From specific marketplace
claude plugin uninstall <name>  # Remove plugin
claude plugin enable <name>     # Enable disabled plugin
claude plugin disable <name>    # Disable plugin
claude plugin update <name>     # Update plugin
claude plugin validate <path>   # Validate manifest
claude plugin marketplace       # Manage marketplaces

claude update - Self-Update

claude update                   # Check and install updates

claude doctor - Diagnostics

claude doctor                   # Check health/issues

claude install - Native Build

claude install                  # Install native build
claude install stable           # Specific version
claude install latest           # Latest version

Slash Commands (Interactive)

Custom Commands

Create .claude/commands/fix-issue.md:

Fix GitHub issue #$ARGUMENTS

1. Read the issue details
2. Identify the problem
3. Implement the fix
4. Write tests
5. Create a commit

Usage: /project:fix-issue 1234

Configuration

User settings (~/.claude/settings.json):

{
  "model": "claude-sonnet-4-5-20250929",
  "verbose": false,
  "theme": "dark"
}

Project settings (.claude/settings.json):

{
  "allowedTools": ["Bash(git:*)", "Read", "Edit"],
  "disallowedTools": ["Bash(rm:*)"]
}

CLI Flags

Core

-p, --print                 Non-interactive mode
-c, --continue              Continue last conversation
-r, --resume <ID>           Resume specific session
-v, --version               Show version

Model & Config

--model <MODEL>             sonnet|opus|haiku or full name
--fallback-model <MODEL>    Fallback when overloaded
--settings <FILE>           Load settings JSON
--setting-sources <LIST>    user,project,local
--session-id <UUID>         Use specific session ID

System Prompt

--system-prompt <TEXT>      Replace default prompt
--append-system-prompt <T>  Append to default
--system-prompt-file <F>    Replace with file (print only)
--append-system-prompt-file Replace with file (print only)

Agent & Tools

--agent <NAME>              Specify agent
--agents <JSON>             Define custom subagents
--tools <LIST>              Restrict built-in tools
--allowedTools <LIST>       Auto-approve tools
--disallowedTools <LIST>    Remove tools from context

Permissions

--permission-mode <MODE>    acceptEdits|bypassPermissions|default|delegate|dontAsk|plan
--dangerously-skip-permissions  Skip all prompts (DANGEROUS)
--allow-dangerously-skip-permissions  Enable bypass option

Output

--output-format <FMT>       text|json|stream-json
--input-format <FMT>        text|stream-json
--include-partial-messages  Include streaming chunks
--verbose                   Verbose logging
--debug [FILTER]            Debug mode with filtering

Advanced

--max-turns <N>             Limit agent turns (print only)
--max-budget-usd <AMT>      Spending limit (print only)
--json-schema <SCHEMA>      Validate JSON output
--chrome / --no-chrome      Chrome integration
--ide                       IDE auto-connect
--fork-session              Create new session on resume
--no-session-persistence    Don't save session
--add-dir <DIRS>            Additional directories
--plugin-dir <DIRS>         Load plugins
--disable-slash-commands    Disable all skills
--mcp-config <FILES>        MCP server configs
--strict-mcp-config         Only use specified MCP
--betas <HEADERS>           Beta API headers

Custom Subagents

claude --agents '{
  "reviewer": {
    "description": "Code reviewer. Use after changes.",
    "prompt": "You are a senior code reviewer...",
    "tools": ["Read", "Grep", "Glob"],
    "model": "sonnet"
  }
}'

Common Patterns & Edge Cases

CI/CD Integration

Codex in GitHub Actions:

- name: Run Codex
  run: |
    echo "${{ secrets.OPENAI_API_KEY }}" | codex login --with-api-key
    codex exec --json -o result.txt "fix linting errors"

Claude in CI:

- name: Run Claude
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
  run: |
    claude -p --output-format json "review this PR" > review.json

Handling Rate Limits

Codex: Automatic backoff built-in.

Claude: Use --fallback-model:

claude -p --fallback-model haiku "quick task"

Working with Large Codebases

# Codex: Use /compact to free context
# In session: /compact

# Claude: Use /compact or start fresh
claude --no-session-persistence -p "analyze src/"

Multi-Directory Access

# Codex
codex --add-dir ../shared-lib --add-dir ../config

# Claude
claude --add-dir ../shared-lib ../config

Structured Output

Codex:

codex exec --output-schema schema.json "generate API spec"

Claude:

claude -p --json-schema '{"type":"object","properties":{"name":{"type":"string"}}}' "extract data"

Image Input

Codex:

codex -i screenshot.png "explain this UI"
codex -i img1.png -i img2.png "compare these"

Claude:

# Via file reference in prompt
claude "analyze the image at ./screenshot.png"

Session Forking

# Codex: /fork in session

# Claude
claude -r "session-id" --fork-session "try alternative approach"

MCP Server Debugging

Codex:

codex mcp list --json | jq .

Claude:

claude --debug "mcp" --mcp-config ./mcp.json

Troubleshooting

Authentication Issues

Sandbox Issues

MCP Server Issues

Performance Issues

Best Practices

1. Start with read-only for exploration, escalate as needed
2. Use sessions - resume work instead of starting fresh
3. Create AGENTS.md/CLAUDE.md for project-specific instructions
4. Leverage MCP servers for external integrations
5. Use structured output in CI/CD for parsing
6. Set spending limits with --max-budget-usd
7. Review diffs before applying (/diff, codex cloud diff)
8. Commit checkpoints before major changes