codex（Codex 编程助手）

Codex Execution Skill

Prerequisites

• Codex CLI installed and configured (~/.codex/config.toml)
• Verify availability: codex --version on first use per session

Workflow Checklist

For every Codex task, follow this sequence:

1. ☐ Detect HPC/Slurm environment:

- Check if running on HPC cluster (look for /home/woody/, /home/hpc/, Slurm env vars) - If HPC detected: Always use --yolo flag to bypass Landlock sandbox restrictions

1. ☐ Ask user for execution parameters via AskUserQuestion (single prompt):

- Model: gpt-5, gpt-5-codex, or default - Reasoning effort: minimal, low, medium, high

1. ☐ Determine sandbox mode based on task:

- read-only: Code review, analysis, documentation - workspace-write: Code modifications, file creation - danger-full-access: System operations, network access - HPC override: Always add --yolo flag (bypasses Landlock restrictions)

1. ☐ Build command with required flags: codex exec [OPTIONS] "PROMPT" Essential flags: HPC command pattern (with --yolo to bypass Landlock): codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check \ "Analyze this code: $(cat /path/to/file.py)" 2>/dev/null Note: --yolo is an alias for --dangerously-bypass-approvals-and-sandbox and is REQUIRED on HPC clusters to avoid Landlock sandbox errors. Do not use --full-auto with --yolo as they are incompatible.

- -m <MODEL> (if overriding default) - -c model_reasoning_effort="<LEVEL>" - -s <SANDBOX_MODE> (skip on HPC) - --skip-git-repo-check (if outside git repo) - -C <DIRECTORY> (if changing workspace) - --full-auto (for non-interactive execution, cannot be used with --yolo)

1. ☐ Execute with stderr suppression:

- Append 2>/dev/null to hide thinking tokens - Remove only if user requests verbose output or debugging

1. ☐ Validate execution:

- Check exit code (0 = success) - Summarize output for user - Report errors with actionable solutions - If Landlock/sandbox errors on HPC: verify --yolo flag was used, retry if missing

1. ☐ Inform about resume capability:

- "Resume this session anytime: codex resume"

Command Patterns

🔥 HPC QUICK TIP: On HPC clusters (e.g., /home/woody/, /home/hpc/), ALWAYS add --yolo flag to avoid Landlock sandbox errors. Example: codex exec --yolo -m gpt-5...

Read-Only Analysis

codex exec -m gpt-5 -c model_reasoning_effort="medium" -s read-only \
  --skip-git-repo-check --full-auto "review @file.py for security issues" 2>/dev/null

Stdin Input (bypasses sandbox file restrictions)

cat file.py | codex exec -m gpt-5 -c model_reasoning_effort="low" \
  --skip-git-repo-check --full-auto - 2>/dev/null

Note: Stdin with - flag may not be supported in all Codex CLI versions.

HPC/Slurm Environment (YOLO Mode - Bypass Landlock)

When running on HPC clusters with Landlock security restrictions, use the --yolo flag:

# Primary solution: --yolo flag bypasses Landlock sandbox
codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check \
  "Analyze this code: $(cat /path/to/file.py)" 2>/dev/null

Alternative: Manual Code Injection (if --yolo is unavailable):

# Capture code content and pass directly in prompt
codex exec -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check --full-auto \
  "Analyze this Python code: $(cat file.py)" 2>/dev/null

Or for large files, use heredoc:

codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check "$(cat <<'ENDCODE'
Analyze the following code comprehensively:

$(cat file.py)

Focus on: architecture, algorithms, multi-GPU optimization, potential bugs, code quality.
ENDCODE
)" 2>/dev/null

Note: --yolo is short for --dangerously-bypass-approvals-and-sandbox and is safe on HPC login nodes where you have limited permissions anyway. Do not combine --yolo with --full-auto as they are incompatible.

Code Modification

codex exec -m gpt-5 -c model_reasoning_effort="high" -s workspace-write \
  --skip-git-repo-check --full-auto "refactor @module.py to async/await" 2>/dev/null

Resume Session

echo "fix the remaining issues" | codex exec --skip-git-repo-check resume --last 2>/dev/null

Cross-Directory Execution

codex exec -C /path/to/project -m gpt-5 -c model_reasoning_effort="medium" \
  -s read-only --skip-git-repo-check --full-auto "analyze architecture" 2>/dev/null

Using Profiles

codex exec --profile production -c model_reasoning_effort="high" \
  --full-auto "optimize performance in @app.py" 2>/dev/null

CLI Reference

Core Flags

Configuration Options

Model Reasoning Effort (-c model_reasoning_effort="<LEVEL>"):

• minimal: Quick tasks, simple queries
• low: Standard operations, routine refactoring
• medium: Complex analysis, architectural decisions (default)
• high: Critical code, security audits, complex algorithms

Model Verbosity (-c model_verbosity="<LEVEL>"):

• low: Minimal output
• medium: Balanced detail (default)
• high: Verbose explanations

Approval Prompts (-c approvals="<WHEN>"):

• on-request: Before any tool use
• on-failure: Only on errors (default for --full-auto)
• untrusted: Minimal prompts
• never: No interruptions (use with caution)

Configuration Management

Config File Location

~/.codex/config.toml

Runtime Overrides

# Override single setting
codex exec -c model="gpt-5" "task"

# Override multiple settings
codex exec -c model="gpt-5" -c model_reasoning_effort="high" "task"

Using Profiles

Define in config.toml:

[profiles.research]
model = "gpt-5"
model_reasoning_effort = "high"
sandbox = "read-only"

[profiles.development]
model = "gpt-5-codex"
sandbox = "workspace-write"

Use with:

codex exec --profile research "analyze codebase"

Resume Behavior

Automatic inheritance:

• Model selection
• Reasoning effort
• Sandbox mode
• Configuration overrides

Resume syntax:

# Resume last session
codex exec resume --last

# Resume with new prompt
codex exec resume --last "continue with next steps"

# Resume via stdin
echo "new instructions" | codex exec resume --last 2>/dev/null

# Resume specific session
codex exec resume <SESSION_ID> "follow-up task"

Flag injection (between exec and resume):

# Change reasoning effort for resumed session
codex exec -c model_reasoning_effort="high" resume --last

Error Handling

Validation Loop

1. Execute command
2. Check exit code (non-zero = failure)
3. Report error with context
4. Ask user for direction via AskUserQuestion
5. Retry with adjustments or escalate

Permission Requests

Before using high-impact flags, request user approval via AskUserQuestion:

• --full-auto: Automated execution
• -s danger-full-access: System-wide access
• --yolo / --dangerously-bypass-approvals-and-sandbox:

- HPC clusters: No approval needed (required for operation) - Personal machines: Request approval (full system access)

Partial Success Handling

When output contains warnings:

1. Summarize successful operations
2. Detail failures with context
3. Use AskUserQuestion to determine next steps
4. Propose specific adjustments

Troubleshooting

File Access Blocked

Symptom: "shell is blocked by the sandbox" or permission errors

Root cause: Sandbox read-only mode restricts file system

Solutions (priority order):

1. Stdin piping (recommended): cat target.py | codex exec -m gpt-5 -c model_reasoning_effort="medium" \ --skip-git-repo-check --full-auto - 2>/dev/null
2. Explicit permissions: codex exec -m gpt-5 -s read-only \ -c 'sandbox_permissions=["disk-full-read-access"]' \ --skip-git-repo-check --full-auto "@file.py" 2>/dev/null
3. Upgrade sandbox: codex exec -m gpt-5 -s workspace-write \ --skip-git-repo-check --full-auto "review @file.py" 2>/dev/null

Invalid Flag Errors

Symptom: "unexpected argument '--add-dir' found"

Cause: Flag does not exist in Codex CLI

Solution: Use -C <DIR> to change directory:

codex exec -C /target/dir -m gpt-5 --skip-git-repo-check \
  --full-auto "task" 2>/dev/null

Exit Code Failures

Symptom: Non-zero exit without clear message

Diagnostic steps:

1. Remove 2>/dev/null to see full stderr
2. Verify installation: codex --version
3. Check configuration: cat ~/.codex/config.toml
4. Test minimal command: codex exec -m gpt-5 "hello world"
5. Verify model access: codex exec --model gpt-5 "test"

Model Unavailable

Symptom: "model not found" or authentication errors

Solutions:

1. Check configured model: grep model ~/.codex/config.toml
2. Verify API access: Ensure valid credentials
3. Try alternative model: -m gpt-5-codex
4. Use OSS fallback: --oss (requires Ollama)

Session Resume Fails

Symptom: Cannot resume previous session

Diagnostic steps:

1. List recent sessions: codex history
2. Verify session ID format
3. Try --last flag instead of specific ID
4. Check if session expired or was cleaned up

HPC/Slurm Sandbox Failures

Symptom: "Landlock sandbox error", "LandlockRestrict", or all file operations fail

Root Cause: HPC clusters use Landlock/seccomp kernel security modules that block Codex's default sandbox

✅ SOLUTION: Use the --yolo flag (priority order):

1. YOLO Flag (PRIMARY SOLUTION - WORKS ON HPC): # Bypasses Landlock restrictions completely codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check \ "Analyze this code: $(cat /full/path/to/file.py)" 2>/dev/null Why this works: --yolo (alias for --dangerously-bypass-approvals-and-sandbox) disables the Codex sandbox entirely, allowing direct file access on HPC systems. Note: Do not use --full-auto with --yolo as they are incompatible.
2. Manual Code Injection (fallback if --yolo unavailable): # Pass code directly in prompt via command substitution codex exec -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check --full-auto \ "Analyze this code comprehensively: $(cat /full/path/to/file.py)" 2>/dev/null
3. Heredoc for Long Code: codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check "$(cat <<'EOF' Analyze the following Python code for architecture, bugs, and optimization opportunities: $(cat /home/user/script.py) Provide technical depth with actionable insights. EOF)" 2>/dev/null
4. Run on Login Node (if compute node blocks outbound): # SSH to login node first, then run codex there (not in Slurm job) ssh login.cluster.edu codex exec --yolo -m gpt-5 --skip-git-repo-check "analyze @file.py" 2>/dev/null
5. Use Apptainer/Singularity (if cluster supports): # Build image with Codex installed, then run via Slurm singularity exec codex.sif codex exec --yolo -m gpt-5 "task"

Best Practice for HPC:

• Always use --yolo flag on HPC clusters - it's safe on login nodes where you already have limited permissions
• Run analysis on login nodes, submit only heavy compute jobs to Slurm
• Keep code files on shared filesystem readable from login nodes
• Combine --yolo with $(cat file.py) for maximum compatibility

Best Practices

Reasoning Effort Selection

• minimal: Syntax fixes, simple renaming
• low: Standard refactoring, basic analysis
• medium: Complex refactoring, architecture review
• high: Security audits, algorithm optimization, critical bugs

Sandbox Mode Selection

• read-only: Default for any analysis or review
• workspace-write: File modifications only
• danger-full-access: Network operations, system commands (rare)

Stderr Suppression

• Always use 2>/dev/null unless:

- User explicitly requests thinking tokens - Debugging failed commands - Troubleshooting configuration issues

Profile Usage

Create profiles for common workflows:

• review: High reasoning, read-only
• refactor: Medium reasoning, workspace-write
• quick: Low reasoning, read-only
• security: High reasoning, workspace-write

Stdin vs File Reference

• Stdin: Single file analysis, avoids permissions
• File reference: Multi-file context, codebase-wide changes

Safety Guidelines

HPC Clusters - --yolo is SAFE and REQUIRED:

• HPC login nodes already have strict permissions (no root access, no network modification)
• --yolo bypasses Codex sandbox but you still operate within HPC user restrictions
• Always use --yolo on HPC to avoid Landlock errors

General Use - Exercise Caution:

• Don't use --yolo on unrestricted systems (your laptop, cloud VMs with full sudo)
• Prefer --full-auto + -s workspace-write for normal development

Always verify before:

• Using danger-full-access sandbox (outside HPC)
• Disabling approval prompts on production systems
• Running with --yolo on personal machines with sudo access

Ask user approval for:

• First-time workspace-write usage
• System-wide access requests
• Destructive operations (deletions, migrations)

Advanced Usage

CI/CD Integration

codex exec --json -o result.txt -m gpt-5 \
  -c model_reasoning_effort="medium" \
  --skip-git-repo-check --full-auto \
  "run security audit on changed files" 2>/dev/null

Batch Processing

for file in *.py; do
  cat "$file" | codex exec -m gpt-5 -c model_reasoning_effort="low" \
    --skip-git-repo-check --full-auto "lint and format" - 2>/dev/null
done

Multi-Step Workflows

# Step 1: Analysis
codex exec -m gpt-5 -c model_reasoning_effort="high" -s read-only \
  --full-auto "analyze @codebase for architectural issues" 2>/dev/null

# Step 2: Resume with changes
echo "implement suggested refactoring" | \
  codex exec -s workspace-write resume --last 2>/dev/null

When to Escalate

If errors persist after troubleshooting:

1. Check documentation: WebFetch https://developers.openai.com/codex/cli/reference WebFetch https://developers.openai.com/codex/local-config#cli
2. Report to user:

- Error message verbatim - Attempted solutions - Configuration details - Exit codes and stderr output

1. Request guidance:

- Alternative approaches - Configuration adjustments - Manual intervention points

Model Selection Guide

Common Workflows

Code Review Workflow

1. Ask user: model + reasoning effort
2. Run read-only analysis
3. Present findings
4. If changes needed: resume with workspace-write
5. Validate changes
6. Inform about resume capability

Refactoring Workflow

1. Ask user: model + reasoning effort
2. Analyze current code (read-only)
3. Propose changes
4. Get user approval
5. Apply changes (workspace-write)
6. Run validation/tests
7. Report results

Security Audit Workflow

1. Use high reasoning effort
2. Run comprehensive analysis (read-only)
3. Document findings
4. Propose fixes
5. Apply fixes if approved (workspace-write)
6. Re-audit to verify
7. Generate report