Codex Image

codex-image — AI Image Generation via Codex OAuth

Generate images using OpenAI's gpt-image-2 model through Codex CLI. No API key required — uses Codex OAuth (ChatGPT login) authentication.

User-facing quoted messages below are bilingual (EN / KO) — print the variant matching the user's chat language.

How it works

User prompt → Claude Code (/codex-image)
  → codex exec (OAuth token auto-managed)
    → built-in image_gen tool (gpt-image-2)
      → ~/.codex/generated_images/<session>/
        → copy to project root

Important: OAuth tokens cannot call OpenAI REST API directly (returns 401). Must go through codex exec which handles auth internally.

Step 1 — Verify Codex CLI & Auth

which codex 2>/dev/null && codex --version 2>/dev/null || echo "NOT_FOUND"

If NOT_FOUND, stop:

"Codex CLI not installed. Run npm install -g @openai/codex then codex login." "Codex CLI 없음. npm install -g @openai/codex 후 codex login 실행해."

codex login status 2>&1

If not "Logged in":

"Codex login required. Run codex login in terminal. OAuth login enables image generation without API key." "Codex 로그인 필요. 터미널에서 codex login 실행. OAuth 로그인하면 API 키 없이 이미지 생성 가능."

Step 2 — Parse Arguments

Extract from $ARGUMENTS:

Remaining text → image prompt.

If prompt is empty, ask via AskUserQuestion:

"What image should I generate? Enter a prompt." "어떤 이미지를 생성할까? 프롬프트를 입력해줘."

Step 2.5 — Background normalization (avoid the transparency checkerboard)

Known gpt-image-2 failure mode. When a prompt asks for a transparent background, gpt-image-2 does NOT return true alpha — it paints a literal gray-and-white checkerboard (the pattern editors use to display transparency) into the RGB pixels. The result looks broken on any real backdrop. This skill therefore never passes a transparency request straight through.

Strip transparency phrasing from the prompt before generating (the Step 4 task carries a hard background rule as a second guard, so this is belt-and-suspenders):

# Neutralize "transparent background" / "transparent bg" / "no background" so the model
# does not bake a checkerboard. Caller-named backdrop colors (e.g. "#FAFAF9 background",
# "white background") are kept and become the clean solid fill.
_PROMPT=$(printf '%s' "${_PROMPT}" \
  | sed -E 's/(fully |truly )?transparent[ -]+(background|bg)/clean solid background/Ig' \
  | sed -E 's/\bno background\b/clean solid background/Ig')

If the user genuinely needs a cut-out asset, tell them gpt-image-2 (via the codex image_gen tool) cannot emit reliable alpha; generate on a flat solid color and remove the background afterward with an image editor. Do NOT request "transparent" to get it.

Step 3 — Determine Save Path

_PROJECT_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
_OUT_DIR="${_PROJECT_ROOT}"
_TIMESTAMP=$(date +%Y%m%d-%H%M%S)
# --filename takes precedence; fall back to timestamp stem
_FILENAME="${_FILENAME_ARG:-codex-image-${_TIMESTAMP}}"

• If --out specified, use that path
• If --filename specified, that stem is used as-is
• Single image (default stem): codex-image-<timestamp>.png
• Single image (--filename foo): foo.png
• Multiple (-n > 1): <stem>-1.png, <stem>-2.png, ...
• Never overwrite existing files (named outputs included — reusing a --filename requires deleting the existing file first or changing --out)

Step 4 — Generate Image

codex exec "Perform the following tasks:
1. Use the built-in image_gen tool to generate an image.
2. Prompt: '${_PROMPT}'
3. Background rule (MUST follow): the image must have a clean, single solid flat background. NEVER paint a transparency/alpha checkerboard — i.e. no alternating gray-and-white squares anywhere. If the prompt implies a 'transparent' or 'no' background, render a clean solid background of the nearest plain color instead (use the backdrop color named in the prompt if any, otherwise plain white).
4. Size: ${_SIZE}
5. Quality: ${_QUALITY}
6. Count: ${_N}
7. Copy the generated image to '${_OUT_DIR}/${_FILENAME}.png'. For multiple images use -1.png, -2.png suffix.
8. Print the saved file path and size." \
  -C "${_PROJECT_ROOT}" \
  -s workspace-write \
  -c 'model_reasoning_effort="medium"' \
  --skip-git-repo-check \
  2>&1

timeout: 120000ms (2 min)

Required flags

• -s workspace-write — file write permission
• --skip-git-repo-check — works outside git repos

Internal flow (Codex side)

1. Codex calls built-in image_gen tool (gpt-image-2)
2. Image saved to ~/.codex/generated_images/<session-id>/ig_*.png
3. Codex copies file to specified project path
4. Reports file path and size

Step 5 — Display Result

═══════════════════════════════════════════════
IMAGE GENERATED / 이미지 생성 완료
═══════════════════════════════════════════════
Prompt: <prompt used>
Size: <size>
Quality: <quality>
Count: <n>
Auth: OAuth (ChatGPT)
───────────────────────────────────────────────
<saved file path(s)>
═══════════════════════════════════════════════

Always display the generated image using the Read tool.

Step 6 — Follow-up

• "Run /codex-image again to generate another image."
• For Next.js projects: suggest moving to public/images/ if needed.
• ppt-master integration: this project's PPT pipeline generates images through image_gen.py --manifest whose default backend codex (scripts/image_backends/backend_codex.py) uses the same codex exec + image_gen mechanism as this skill — no API key or .env needed, only codex login. Use /codex-image directly for one-off images outside the pipeline (e.g. re-rolling a single asset with --out <project>/images --filename <slot>).

Error Handling

Rules

• Always use the Read tool to display generated images
• Never overwrite existing files — always use timestamped filenames
• OAuth only — do not attempt direct REST API calls with OAuth token (returns 401)
• Verify prompt intent before generating