Codex CLI Skill

Use the codex CLI (OpenAI GPT-5.5 + GPT-5.4 Image 2) as a complementary tool to Claude. Codex's strengths are careful file reading, edge-case spotting, role-play fidelity, and unlimited high-quality image generation under subscription auth.

When To Use Codex (Decision Tree)

User intent
├── "Review/audit my code/PR/diff"        → Codex review  (see references/code-review-workflows.md)
├── "Audit/review my plan or spec"        → Codex plan audit (see references/plan-audit-workflows.md)
├── "Generate an image/banner/illustration/icon/infographic/visual asset" → Codex image gen (see references/image-generation.md)
├── "Get a second opinion from Codex"     → Codex exec one-shot
├── "Pretend to be Gemini/Qwen/o3"        → Codex role-play (see references/best-practices.md)
└── "Just run codex on X programmatically" → Codex exec (see references/cli-reference.md)

Do NOT use Codex when:

• The task is small, mechanical, or fully solvable by Claude in the current session (waste of latency/credits).
• The user wants a final implementation pushed to disk by Codex — prefer Claude for in-session edits, use Codex for analysis only unless explicitly asked.
• Sensitive data is in the working dir without user clearance (Codex sends content to OpenAI).

Why Codex Beats Claude For Certain Tasks

Core Invocation Patterns

All Codex calls are non-interactive (codex exec or codex review). NEVER call bare codex in a script — it opens the TUI.

Pattern 1: One-shot prompt

codex exec -C "$PWD" --skip-git-repo-check </dev/null \
  "Audit ./plans/foo.md for missing edge cases"

Pattern 2: Full-auto write (convenience for automation)

codex exec --full-auto -C "$PWD" </dev/null "Generate hero.png in ./assets"
# --full-auto == --sandbox workspace-write + --approval-policy on-failure

Pattern 3: Code review against branch

codex review --base main "Focus on auth and data-loss risks"

Pattern 4: Capture last message only (for piping into Claude)

codex exec -o /tmp/codex-out.md "Review this diff" -C "$PWD"

Pattern 5: JSON event stream (programmatic parsing)

codex exec --json "..." | jq -r 'select(.type=="agent_message") | .message'

Critical Defaults & Avoidances

• Always pass `-C "$PWD"` when invoking from a script so Codex roots at the right project, not ~/.codex.
• Pass `--skip-git-repo-check` when CWD is not a git repo, otherwise Codex aborts with Not inside a trusted directory.
• Redirect stdin (</dev/null) in non-interactive scripts — without it, Codex hangs reading stdin when both arg-prompt and a TTY-less stdin are present.
• Default sandbox is good (workspace-write). Use read-only for pure review/audit. Use danger-full-access ONLY when user explicitly authorizes.
• Network is blocked by default inside workspace-write / read-only. Enable per-call: -c 'sandbox_workspace_write.network_access=true'.
• Approval-policy quirk: setting policy via config can silently revert to on-failure (open bug #11885). Pass --approval-policy directly on the CLI.
• NEVER use `--dangerously-bypass-approvals-and-sandbox` unless the parent process is already externally sandboxed and the user asked for it.
• Prefer `$imagegen` skill invocation for images — the literal token $imagegen in the prompt loads Codex's bundled official imagegen skill (correct model routing, save-path policy, transparency flow). scripts/codex-generate-image.sh auto-prepends it.
• Images land in two places — ~/.codex/generated_images/<session>/ig_*.png (built-in tool) OR CWD (if prompt says "Save as X"). scripts/codex-generate-image.sh snapshots both.
• `gpt-image-2` does NOT support native transparent PNG — official chroma-key workflow: generate on #00ff00 background then run scripts/codex-strip-chroma-key.sh. Only escalate to gpt-image-1.5 --background transparent after asking user (upstream rule: never silently downgrade).
• `gpt-image-2` ignores exact pixel dimensions — verified empirically (a "256x256" request returned 1254x1254). Resize via ImageMagick if exact dimensions matter.
• Don't double-instruct: Codex respects AGENTS.md cascading from project root → CWD. Don't repeat global rules in the prompt.
• Don't pipe huge files: prefer -i path for images; let Codex read text files via its own tools. Stdin >100KB is slow/truncatable.
• Codex sessions persist in ~/.codex/sessions/. Use --ephemeral for one-off prompts you don't want logged.
• Watch the dual-window rate limit (subscription): 5-hour window AND weekly window both must have budget. Check via /status.
• Platform sandbox gotchas (see references/best-practices.md): Windows error 1385, macOS MCP shell-exec blocked in workspace-write, Linux symlinked .codex/ panics.

Workflows

Output Handling

After every Codex call:

1. Read the output — never assume Codex succeeded silently.
2. Surface findings to user verbatim when Codex audits something; do not summarize away edge cases (that defeats the point of asking Codex).
3. If Codex generated images, run the post-move logic in scripts/codex-generate-image.sh or manually:

   latest_session=$(ls -t ~/.codex/generated_images/ | head -1)
   mv ~/.codex/generated_images/"$latest_session"/* ./assets/ && rm -rf ~/.codex/generated_images/"$latest_session"

1. Cite findings with file:line when relaying Codex's review back to user.

Security Policy

This skill handles: invoking codex CLI, parsing its stdout, moving generated images, optionally piping diffs/plans into Codex.

This skill does NOT handle: storing OpenAI API keys (uses existing ~/.codex/auth.json), modifying Codex config, executing arbitrary user code outside codex exec.

Refuse to: bypass sandbox without explicit user authorization, send .env or credential files to Codex, run codex exec against directories the user didn't ask you to touch.