Files
gh-lucklyric-cc-dev-tools-p…/skills/codex/references/advanced-patterns.md
2025-11-30 08:38:18 +08:00

11 KiB

Advanced Configuration Examples


⚠️ CRITICAL: Always Use codex exec

ALL commands in this document use codex exec - this is mandatory in Claude Code.

NEVER: codex -m ... or codex --flag ... (will fail with "stdout is not a terminal") ALWAYS: codex exec -m ... or codex exec --flag ... (correct non-interactive mode)

Claude Code's bash environment is non-terminal. Plain codex commands will NOT work.


Custom Model Selection

Example 1: Force GPT-5 for Code Task

User Request: "Use GPT-5.1 (not Codex) to review this code for architecture issues"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  "Review this code for architecture issues"

Why: Even though it's code-related, user wants architectural review (high-reasoning) rather than code editing.


Example 2: Explicit GPT-5-Codex for Implementation

User Request: "Use GPT-5.1-Codex to implement the authentication module"

Skill Executes:

codex exec -m gpt-5.1-codex -s workspace-write \
  -c model_reasoning_effort=high \
  "Implement the authentication module"

Why: Implementation requires file writing and code generation (gpt-5.1-codex specialty).


Workspace Write Permission

Example 3: Allow File Modifications

User Request: "Have Codex refactor this codebase (allow file writing)"

Skill Executes:

codex exec -m gpt-5.1-codex -s workspace-write \
  -c model_reasoning_effort=high \
  "Refactor this codebase for better maintainability"

Permission: workspace-write allows Codex to modify files directly.

⚠️ Warning: Only use workspace-write when you trust the operation and want file modifications.


Example 4: Read-Only Code Review

User Request: "Review this code for security vulnerabilities (read-only)"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  "Review this code for security vulnerabilities"

Permission: read-only prevents file modifications - safer for review tasks.


Web Search Integration

Example 5: Research Latest Patterns

User Request: "Research latest Python async patterns and implement them (enable web search)"

Skill Executes:

codex exec -m gpt-5.1-codex -s workspace-write \
  -c model_reasoning_effort=high \
  --search \
  "Research latest Python async patterns and implement them"

Feature: --search flag enables web search for up-to-date information.


Example 6: Security Best Practices Research

User Request: "Use web search to find latest JWT security best practices, then review this auth code"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  --search \
  "Find latest JWT security best practices and review this auth code"

Reasoning Effort Control

Example 7: Maximum Reasoning for Complex Algorithm

User Request: "Design an optimal algorithm for distributed consensus (maximum reasoning)"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  "Design an optimal algorithm for distributed consensus"

Default: Already uses high reasoning effort.


Example 8: Quick Code Review (Lower Reasoning)

User Request: "Quick syntax check on this code (low reasoning)"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=low \
  "Quick syntax check on this code"

Use Case: Fast turnaround for simple tasks.


Verbosity Control

Example 9: Detailed Explanation

User Request: "Explain this algorithm in detail (high verbosity)"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  -c model_verbosity=high \
  "Explain this algorithm in detail"

Output: Comprehensive, detailed explanation.


Example 10: Concise Summary

User Request: "Briefly review this code (low verbosity)"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  -c model_verbosity=low \
  "Review this code"

Output: Concise, focused feedback.


Working Directory Control

Example 11: Specific Project Directory

User Request: "Work in the backend directory and review the API code"

Skill Executes:

codex exec -m gpt-5.1 -s read-only \
  -c model_reasoning_effort=high \
  -C ./backend \
  "Review the API code"

Feature: -C flag sets working directory for Codex.


Approval Policy

Example 12: Request Approval for Shell Commands

User Request: "Implement the build script (ask before running commands)"

Skill Executes:

codex exec -m gpt-5.1-codex -s workspace-write \
  -c model_reasoning_effort=high \
  -a on-request \
  "Implement the build script"

Safety: -a on-request requires approval before executing shell commands.


Combined Advanced Configuration

User Request: "Use web search to find latest security practices, review my auth module in detail with high reasoning, allow file fixes if needed (ask for approval)"

Skill Executes:

codex exec -m gpt-5.1-codex -s workspace-write \
  -c model_reasoning_effort=high \
  -c model_verbosity=high \
  -a on-request \
  --search \
  "Find latest security practices, review my auth module in detail, and fix issues"

Features:

  • Web search enabled (--search)
  • High reasoning (model_reasoning_effort=high)
  • Detailed output (model_verbosity=high)
  • File writing allowed (workspace-write)
  • Requires approval for commands (-a on-request)

Decision Tree: When to Use GPT-5.1 vs GPT-5.1-Codex

Use GPT-5.1 For:

┌─────────────────────────────────────┐
│     Architecture & Design           │
│  - System architecture              │
│  - API design                       │
│  - Data structure design            │
│  - Algorithm analysis               │
├─────────────────────────────────────┤
│     Analysis & Review               │
│  - Code reviews                     │
│  - Security audits                  │
│  - Performance analysis             │
│  - Quality assessment               │
├─────────────────────────────────────┤
│     Explanation & Learning          │
│  - Concept explanations             │
│  - Documentation review             │
│  - Trade-off analysis               │
│  - Best practices guidance          │
└─────────────────────────────────────┘

Use GPT-5.1-Codex For:

┌─────────────────────────────────────┐
│     Code Editing                    │
│  - Modify existing files            │
│  - Implement features               │
│  - Refactoring                      │
│  - Bug fixes                        │
├─────────────────────────────────────┤
│     Code Generation                 │
│  - Write new code                   │
│  - Generate boilerplate             │
│  - Create test files                │
│  - Scaffold projects                │
├─────────────────────────────────────┤
│     File Operations                 │
│  - Multi-file changes               │
│  - Batch updates                    │
│  - Migration scripts                │
│  - Build configurations             │
└─────────────────────────────────────┘

Sandbox Mode Decision Matrix

Task Recommended Sandbox Rationale
Code review read-only No modifications needed
Architecture design read-only Planning phase only
Security audit read-only Analysis without changes
Implement feature workspace-write Requires file modifications
Refactor code workspace-write Must edit existing files
Generate new files workspace-write Creates new files
Bug fix workspace-write Edits source files

Configuration Profiles

Create a Config Profile

You can create reusable configuration profiles in ~/.codex/config.toml:

[profiles.review]
model = "gpt-5.1"
sandbox = "read-only"
model_reasoning_effort = "high"
model_verbosity = "medium"

[profiles.implement]
model = "gpt-5.1-codex"
sandbox = "workspace-write"
model_reasoning_effort = "high"
approval_policy = "on-request"

Use Profile in Skill

User Request: "Use the review profile to analyze this code"

Skill Executes:

codex -p review "Analyze this code"

Result: Uses all settings from [profiles.review].


Best Practices

1. Match Model to Task Type

  • Thinking/Design → GPT-5.1
  • Doing/Coding → GPT-5.1-Codex

2. Use Safe Defaults, Override Intentionally

  • Default to read-only unless file writing is explicitly needed
  • Default to high reasoning for complex tasks
  • Reduce reasoning effort only for simple, quick tasks

3. Combine Web Search with High Reasoning

For best results researching current practices:

codex exec -m gpt-5.1--search \
  -c model_reasoning_effort=high \
  "Research latest distributed systems patterns"

4. Request Approval for Risky Operations

Use -a on-request when:

  • Working with production code
  • Running shell commands
  • Making broad changes

Common Patterns

Pattern 1: Research → Design → Implement

Phase 1 - Research (GPT-5.1 + web search):

codex exec -m gpt-5.1--search \
  -c model_reasoning_effort=high \
  "Research latest authentication patterns"

Phase 2 - Design (GPT-5.1 + high reasoning):

codex exec resume --last
# "Design the authentication system based on research"

Phase 3 - Implement (GPT-5.1-Codex + workspace-write):

codex exec -m gpt-5.1-codex -s workspace-write \
  -c model_reasoning_effort=high \
  "Implement the authentication system we designed"

Pattern 2: Review → Fix → Verify

Review (GPT-5.1 + read-only):

codex exec -m gpt-5.1 -s read-only \
  "Review this code for security issues"

Fix (GPT-5.1-Codex + workspace-write):

codex exec resume --last
# "Fix the security issues identified"

Verify (GPT-5.1 + read-only):

codex exec resume --last
# "Verify the fixes are correct"

Next Steps