codex（Codex 编程助手）

Codex: High-Reasoning AI Assistant for Claude Code

DEFAULT MODEL: GPT-5.4 with Read-Only Default

GPT-5.4 is the default model for ALL tasks. Sandbox is read-only by default - only use workspace-write when user explicitly requests file editing.

• gpt-5.4: OpenAI's most capable frontier model - unified for both code and general tasks
• gpt-5.4-fast: Faster variant for speed-sensitive tasks (use ONLY when user explicitly requests "fast", "quick", or "speed")
• Sandbox default: Always read-only unless user explicitly requests editing
• Explicit editing: Only when user says "edit", "modify", "write changes", etc., use workspace-write
• Always use -c model_reasoning_effort=xhigh for maximum capability

# Default (read-only)
codex exec -m gpt-5.4 -s read-only \
  -c model_reasoning_effort=xhigh \
  "analyze this function implementation"

# With explicit edit request
codex exec -m gpt-5.4 -s workspace-write \
  -c model_reasoning_effort=xhigh \
  "edit this file to add the feature"

# Fast mode (on demand only)
codex exec -m gpt-5.4-fast -s read-only \
  "quick analysis of this function"

Model Fallback Chain

If the primary model is unavailable, fallback gracefully:

1. gpt-5.4 → gpt-5.4-fast → gpt-5.4
2. Reasoning effort: xhigh → high → medium

CRITICAL: Always Use codex exec

MUST USE: codex exec for ALL Codex CLI invocations in Claude Code.

NEVER USE: codex (interactive mode) - will fail with "stdout is not a terminal" ALWAYS USE: codex exec (non-interactive mode)

Examples:

• codex exec -m gpt-5.4 "prompt" (CORRECT)
• codex -m gpt-5.4 "prompt" (WRONG - will fail)
• codex exec resume --last (CORRECT)
• codex resume --last (WRONG - will fail)

Why? Claude Code's bash environment is non-terminal/non-interactive. Only codex exec works in this environment.

IMPORTANT: Interactive vs Exec Mode Flags

Some Codex CLI flags are ONLY available in interactive mode, NOT in codex exec.

⚠️ Web Search Note (v0.114.0+): The web_search_request feature flag is deprecated. Web search is now built-in when the model supports it. No --enable flag is needed in exec mode.

For approval control in exec mode:

# CORRECT - works in codex exec
codex exec --full-auto "task"
codex exec -c approval_policy=on-request "task"

# WRONG - -a only works in interactive mode
codex -a on-request "task"

Trigger Examples

This skill activates when users say phrases like:

• "Use codex to analyze this architecture"
• "Ask codex about this design decision"
• "Run codex on this problem"
• "Call codex for help with this implementation"
• "I need GPT-5 reasoning for this task"
• "Get OpenAI's high-reasoning model on this"
• "Continue with codex" or "Resume the codex session"
• "Codex, help me with..." or simply "Codex"

When to Use This Skill

This skill should be invoked when:

• User explicitly mentions "Codex" or requests Codex assistance
• User needs help with complex coding tasks, algorithms, or architecture
• User requests "high reasoning" or "advanced implementation" help
• User needs complex problem-solving or architectural design
• User wants to continue a previous Codex conversation

How It Works

Detecting New Codex Requests

When a user makes a request, determine sandbox based on explicit edit request:

Step 1: Model Selection

• Default: gpt-5.4 for ALL tasks (code and general)
• Fast mode: gpt-5.4-fast ONLY when user explicitly requests speed

Step 2: Determine Sandbox (Edit Permission)

• Default: read-only - safe for all tasks unless user explicitly requests editing
• Explicit edit request: workspace-write - ONLY when user explicitly says to edit/modify/write files

Read-only examples: "Analyze this function", "Design a queue", "Explain this algorithm" Edit examples: "Edit this file to fix the bug", "Modify the function", "Update the README"

⚠️ Important: Use workspace-write ONLY when user says "edit", "modify", "write changes", "save", etc.

Bash CLI Command Structure

See the DEFAULT MODEL section above for complete command templates. Key points:

• Always use codex exec (non-interactive mode required)
• Web search is built-in (no flag needed as of v0.114.0)
• See references/command-patterns.md for additional patterns

Model Selection Logic

Model: gpt-5.4 for ALL tasks (unified model, no task-based selection needed)

• gpt-5.4-fast ONLY when user explicitly requests speed/fast mode

Sandbox:

• read-only (DEFAULT): Analysis, review, explanation, any task without explicit edit request
• workspace-write: ONLY when user explicitly says "edit", "modify", "write changes", "save"

Fallback: gpt-5.4 → gpt-5.4-fast → gpt-5.4. See fallback chain in DEFAULT MODEL section.

Default Configuration

All Codex invocations use these defaults unless user specifies otherwise:

CLI Flags Reference

Codex CLI Version: 0.114.0+

See: references/cli-features.md for the complete CLI flags table and feature documentation.

Key flags for this skill:

• -m, --model - Model selection (gpt-5.4, gpt-5.4-fast)
• -s, --sandbox - Sandbox mode (read-only, workspace-write)
• -c, --config - Config overrides (e.g., model_reasoning_effort=xhigh)
• --enable / --disable - Feature toggles (e.g., multi_agent)

Configuration Parameters

Pass these as -c key=value:

• model_reasoning_effort: none, minimal, low, medium, high, xhigh

- CLI default: high - The Codex CLI defaults to high reasoning - Skill default: xhigh - This skill explicitly uses xhigh for maximum capability - xhigh: Extra-high reasoning for maximum capability - Use xhigh for complex architectural refactoring, long-horizon tasks, or when quality is more important than speed

• model_verbosity: low, medium, high (default: medium)
• model_reasoning_summary: auto, concise, detailed, none (default: auto)
• sandbox_workspace_write.writable_roots: JSON array of additional writable directories (e.g., ["/path1","/path2"])
• approval_policy: untrusted, on-failure, on-request, never (approval behavior)

Additional Writable Directories:

Use --add-dir flag (preferred) or config:

# Using --add-dir for multiple directories
codex exec --add-dir /path1 --add-dir /path2 "task"

# Alternative - config approach
codex exec -c 'sandbox_workspace_write.writable_roots=["/path1","/path2"]' "task"

Model Selection Guide

Available Models:

• gpt-5.4 - ALL tasks (default, highest capability)
• gpt-5.4-fast - Speed-sensitive tasks (on demand only)

Default: gpt-5.4 with xhigh reasoning effort for all tasks.

Session Continuation

Detecting Continuation Requests

When user indicates they want to continue a previous Codex conversation:

• Keywords: "continue", "resume", "keep going", "add to that"
• Follow-up context referencing previous Codex work
• Explicit request like "continue where we left off"

Resuming Sessions

For continuation requests, use the codex resume command:

Resume Most Recent Session (Recommended)

codex exec resume --last

This automatically continues the most recent Codex session with all previous context maintained.

Resume Specific Session

codex exec resume <session-id>

Resume a specific session by providing its UUID. Get session IDs from previous Codex output or by running codex exec resume --last to see the most recent session.

Note: The interactive session picker (codex resume without arguments) is NOT available in non-interactive/Claude Code environments. Always use --last or provide explicit session ID.

Forking Sessions (Interactive Only)

The codex fork command creates a new session from a previous one, allowing exploration of different directions without affecting the original session.

# Fork the most recent session (interactive terminal only)
codex fork --last

# Fork a specific session by ID (interactive terminal only)
codex fork <session-id>

⚠️ Important: codex fork is an interactive-only command. It is NOT available under codex exec and will fail with "stdin is not a terminal" in Claude Code's non-interactive environment.

Workaround for Claude Code: To achieve similar functionality, use codex exec resume --last with a prompt that indicates you want to explore an alternative approach. The session history will be preserved.

Note: Unlike resume which continues the same session, fork creates a new independent session with the same history as a starting point.

Decision Logic: New vs. Continue

Use codex exec -m... "<prompt>" when:

• User makes a new, independent request
• No reference to previous Codex work
• User explicitly wants a "fresh" or "new" session

Use codex exec resume --last when:

• User indicates continuation ("continue", "resume", "add to that")
• Follow-up question building on previous Codex conversation
• Iterative development on same task
• User wants to explore alternatives (provide new direction in prompt)

Session History Management

• Codex CLI automatically saves session history
• No manual session ID tracking needed
• Sessions persist across Claude Code restarts
• Use codex exec resume --last to access most recent session
• Use codex exec resume <session-id> for specific sessions

Error Handling

Simple Error Response Strategy

When errors occur, return clear, actionable messages without complex diagnostics:

Error Message Format:

Error: [Clear description of what went wrong]

To fix: [Concrete remediation action]

[Optional: Specific command example]

Common Errors

Command Not Found

Error: Codex CLI not found

To fix: Install Codex CLI and ensure it's available in your PATH

Check installation: codex --version

Authentication Required

Error: Not authenticated with Codex

To fix: Run 'codex login' to authenticate

After authentication, try your request again.

Invalid Configuration

Error: Invalid model specified

To fix:
- Use 'gpt-5.4' for all tasks
- Use 'gpt-5.4-fast' for speed-sensitive tasks

Example: codex exec -m gpt-5.4 -s workspace-write -c model_reasoning_effort=xhigh "implement feature"
Example (fast): codex exec -m gpt-5.4-fast -s read-only "quick analysis"

Troubleshooting

First Steps for Any Issues:

1. Check Codex CLI built-in help: codex --help, codex exec --help, codex exec resume --help
2. Consult official documentation: https://github.com/openai/codex/tree/main/docs
3. Verify skill resources in references/ directory

Note: Commands like codex --help, codex --version, codex login, and codex logout work without the exec subcommand. The exec requirement only applies to task execution.

Skill not being invoked?

• Check that request matches trigger keywords (Codex, complex coding, high reasoning, etc.)
• Explicitly mention "Codex" in your request
• Try: "Use Codex to help me with..."

Session not resuming?

• Verify you have a previous Codex session (check command output for session IDs)
• Try: codex exec resume --last to resume most recent session
• If no history exists, start a new session first

"stdout is not a terminal" error?

• Always use codex exec instead of plain codex in Claude Code
• Claude Code's bash environment is non-interactive/non-terminal

Errors during execution?

• Codex CLI errors are passed through directly
• Check Codex CLI logs for detailed diagnostics
• Verify working directory permissions if using workspace-write
• Check official Codex docs for latest updates and known issues

Examples

Code Analysis (Read-Only)

codex exec -m gpt-5.4 -s read-only \
  -c model_reasoning_effort=xhigh \
  "Analyze this function implementation"

Code Editing (Explicit Request)

codex exec -m gpt-5.4 -s workspace-write \
  -c model_reasoning_effort=xhigh \
  "Edit this file to implement the feature"

Session Continuation

codex exec resume --last

See: references/examples.md for more examples including web search, file context, and code review patterns.

Code Review Subcommand (v0.71.0+)

The codex review subcommand provides non-interactive code review capabilities:

# Review uncommitted changes (staged, unstaged, untracked)
codex review --uncommitted

# Review changes against a base branch
codex review --base main

# Review a specific commit
codex review --commit abc123

# Review with custom instructions
codex review --uncommitted "Focus on security vulnerabilities"

# Non-interactive via exec
codex exec review --uncommitted

Review Options:

Apply Command (v0.98.0+)

The codex apply command applies the latest diff produced by the Codex agent as a git apply to your local working tree:

# Apply the latest diff from Codex
codex apply

This is useful when Codex generates code changes in read-only mode and you want to apply those changes to your local files.

CLI Features Reference

For detailed CLI feature documentation, see references/cli-features.md.

Quick Reference - Common features:

• Web search is built-in (no flag needed as of v0.114.0)
• -i, --image - Attach images to prompts
• --add-dir - Add writable directories
• --full-auto - Low-friction workspace-write mode
• --json - JSONL output for programmatic processing

File Context Passing

IMPORTANT: Pass file paths to Codex CLI instead of embedding file content in prompts. This enables Codex to read files autonomously.

Quick reference:

• Use -C /path to set working directory
• Use --add-dir /path for additional directories
• Use @path/to/file syntax for explicit file references

# Example: analyze file with explicit @ syntax
codex exec -m gpt-5.4 -s read-only \
  "Analyze @src/auth.ts and compare with @src/session.ts"

# Example: multi-directory analysis
codex exec -m gpt-5.4 -s read-only \
  --add-dir /shared/libs \
  "Review how auth module uses shared utilities"

See: references/file-context.md for complete file context documentation.

Best Practices

1. Use Descriptive Requests

Good: "Help me implement a thread-safe queue with priority support in Python" Vague: "Code help"

Clear, specific requests get better results from high-reasoning models.

2. Indicate Continuation Clearly

Good: "Continue with that queue implementation - add unit tests" Unclear: "Add tests" (might start new session)

Explicit continuation keywords help the skill choose the right command.

3. Specify Permissions When Needed

Good: "Refactor this code (allow file writing)" Risky: Assuming permissions without specifying

Make your intent clear when you need workspace-write permissions.

4. Leverage High Reasoning

The skill defaults to high reasoning effort - perfect for:

• Complex algorithms
• Architecture design
• Performance optimization
• Security reviews

Reference Documentation

For detailed information, consult these reference files:

Core References

• references/file-context.md - File and directory context passing guide
• references/examples.md - Complete command examples by use case
• references/cli-features.md - Feature flags and CLI options

Workflow References

• references/command-patterns.md - Common codex exec usage patterns
• references/session-workflows.md - Session continuation and resume workflows
• references/advanced-patterns.md - Complex configuration and flag combinations

CLI References

• references/codex-help.md - Codex CLI command reference
• references/codex-config.md - Full configuration options