codex

Codex CLI Skill

Codex CLI is OpenAI's terminal-based coding agent. This skill delegates tasks to it via codex exec (non-interactive mode), where Claude constructs the command and summarizes the result.

Attribution: This skill builds upon the original codex skill by @davila7. Key additions: updated model table (GPT-5.3-codex), streamlined interaction (no AskUserQuestion overhead), additional CLI flags documentation, and bilingual trigger support.

When To Use

• Second opinion: User wants a different model's take on code quality, architecture, or bugs
• Codex-specific features: Sandboxed execution, session resume/fork, structured JSON output
• User explicitly asks: "codex", "use codex", "resume codex session"

Running a Task

1. Pick the sandbox mode the task requires (read-only is the safe default):

Task	Sandbox	Command Pattern
Analysis / review	read-only	codex exec --skip-git-repo-check -s read-only "prompt" 2>/dev/null
Apply edits	workspace-write	codex exec --skip-git-repo-check --full-auto "prompt" 2>/dev/null
Network / broad access	danger-full-access	codex exec --skip-git-repo-check -s danger-full-access --full-auto "prompt" 2>/dev/null

2. Always include --skip-git-repo-check — Codex normally requires a git repo, but we're running it from within Claude Code's working directory where this check can fail.

3. Always append 2>/dev/null — Codex streams thinking/progress tokens to stderr. Suppressing stderr keeps Claude's output clean. Only show stderr when debugging.

4. Use --full-auto when the task involves edits — it enables workspace-write sandbox with automatic approval for file changes. Without it, Codex pauses for approval that nobody can give in non-interactive mode.

5. Default model: gpt-5.3-codex with xhigh reasoning effort. Override with -m <model> or --config model_reasoning_effort="<level>".

Complete Example

codex exec --skip-git-repo-check \
  -m gpt-5.3-codex \
  --config model_reasoning_effort="xhigh" \
  -s read-only \
  "Review code/R/processing.R for potential bugs and edge cases" \
  2>/dev/null

Changing Working Directory

Use -C <DIR> to run Codex in a different directory:

codex exec --skip-git-repo-check -C /path/to/project --full-auto "fix the failing tests" 2>/dev/null

Resuming a Session

Codex persists sessions. To continue a previous one:

# Pass prompt as argument (preferred):
codex exec --skip-git-repo-check resume --last "now fix the edge cases you found" 2>/dev/null

# Or pipe via stdin:
echo "now fix the edge cases" | codex exec --skip-git-repo-check resume --last 2>/dev/null

• --last resumes the most recent session in the current directory. Add --all to search across directories.
• Resumed sessions inherit their original model, sandbox, and reasoning effort — do not pass config flags unless the user explicitly wants to override.
• All flags go between exec and resume.
• After Codex completes, remind the user they can say "codex resume" to continue.

Models

Model ID	Role	API Pricing (input/output per 1M tokens)
gpt-5.3-codex	Flagship — best agentic coding	$1.75 / $14.00
gpt-5.3-codex-spark	Near-instant realtime (Pro only)	—
gpt-5.2-codex	Previous gen, still capable	$1.75 / $14.00
gpt-5.1-codex-max	Extended reasoning	$1.25 / $10.00
gpt-5.1-codex	Older flagship	$1.25 / $10.00

All models support 400K context. Cached input gets 90% discount ($0.175/M for 5.3-codex).

Reasoning effort: --config model_reasoning_effort="<level>"

• xhigh — deep analysis (default)
• high — refactoring, architecture
• medium — standard feature work
• low — formatting, docs, quick fixes

Useful Flags

Flag	Purpose
--json	JSONL event stream — useful for parsing Codex output programmatically
-o <path>	Write final assistant message to a file
--output-schema <path>	Constrain response to a JSON Schema
-i <image>	Attach image(s) to prompt
--search	Enable live web search
--ephemeral	Don't persist session (one-off tasks)
--add-dir <path>	Grant write access to additional directories beyond workspace

Error Handling

• If codex exec exits non-zero, report the error and ask for direction — do not retry automatically.
• Before using --full-auto or danger-full-access, confirm with the user unless they already granted permission.
• If output contains warnings or partial results, summarize and ask how to proceed.