Codex Image Gen

Codex Image Gen

Overview

Routes image generation through the user's local OAuth-authenticated Codex CLI (codex exec, non-interactive) instead of any cloud API. The bundled script scripts/codex_generate_image.sh performs the full pipeline (dispatch, locate, resize, verify, archive prompt). Do NOT hand-craft codex exec calls unless the script cannot be used.

Hard rules

• Never report success to the user before every item in the verification checklist has passed.
• Never ask the user for an OpenAI API key; this skill is OAuth-only. If auth is missing or expired, direct the user to run codex login.
• Never hand-edit the produced PNG; if the user wants changes, regenerate with a refined prompt.
• Never state or imply image dimensions without verifying them (the script verifies via sips/ImageMagick).

When not to use this skill

Routing triggers live in the frontmatter description. Skip only when: editing existing SVG/vector/code-native assets; the user explicitly wants the OpenAI API or a non-Codex provider; Codex CLI/OAuth login is missing; or the task is better served by writing HTML/CSS/canvas directly.

Quick Start

Use the bundled script (path is relative to the skill's install directory). It prints the final image path on stdout and exits 0 on success.

scripts/codex_generate_image.sh \
  --prompt  "a sandy beach with a tropical island in the distance and many people enjoying the shore" \
  --output  /tmp/beach.png \
  --size    1024x1024 \
  --style   photorealistic-natural

Exit codes: 0 = success · 1 = no codex/OAuth · 2 = no image produced · 3 = size mismatch after resize · 4 = bad arguments.

Output convention

Write images to an assets/ folder inside the user's workspace, named <task-slug>-<YYYYMMDD-HHMMSS>.png (slug derived from the request). The script saves the full prompt as a sibling .md with the same basename (disable with --no-archive), so every image is reproducible.

Workflow

1. Decide the output path per the convention above: pass assets/<task-slug>-<YYYYMMDD-HHMMSS>.png via --output, preferring the workspace assets/ over /tmp so the artifact persists (the script mkdir -ps parents).
2. Draft a specific, self-contained prompt. Include the subject, scene, composition, style, and any constraints. Translate Chinese prompts to English when targeting photorealistic styles unless the user specifies a language requirement.
3. Confirm before generating — but only when it matters. Text edits are free; each generation is a real Codex call. If the request is vague or a high-impact field is undecided (subject, style slug, text in the image, size/purpose), show the draft and ask option-based questions via the host's structured user-input tool (AskUserQuestion-style) or numbered plain text, always offering "generate as drafted". If the prompt is already specific, skip confirmation. Full protocol: references/prompting.md.
4. Choose a target size. Default is 1024x1024. Codex's built-in image_gen does not guarantee exact pixel dimensions, so the script enforces them after the fact via sips (macOS) with an ImageMagick fallback elsewhere.
5. Optionally pass --style using one of Codex's supported use-case slugs (see references/prompting.md). If omitted, Codex picks a sensible default.
6. Invoke the script with the arguments above. Capture the path it prints on stdout. Note that the archived .md prompt file lands next to the image (same basename).
7. Verify the file exists at the returned path and present the image to the user using whatever display mechanism the host provides.
8. If the user requested edits or variants, run the script again with a refined prompt; do not edit the PNG by hand.

Common pitfalls

• "image_generation feature not enabled" — the script passes --enable image_generation per call; if your Codex version rejects the flag, enable image_generation in ~/.codex/config.toml instead.
• OAuth expired — re-run codex login (NOT codex login --api-key). Do NOT ask the user for an API key.
• Wrong dimensions — Codex may produce a non-matching size; the script resizes via sips (ImageMagick fallback). Expected behavior. If neither tool exists, it warns and keeps the native size.
• Sandbox permissions — broad permissions are scoped per invocation (--sandbox danger-full-access + -c approval_policy="never"); no global config change needed. Pass --sandbox workspace-write to tighten, accepting that generation may fail.

Verification checklist

Confirm all of the following before reporting success to the user:

• Exit code is 0
• File exists at the path printed on stdout
• File is a PNG (file <path> reports PNG image data)
• Pixel dimensions match --size (sips -g pixelWidth -g pixelHeight, or ImageMagick's identify where sips is unavailable)
• Visual content matches the prompt (read the image and confirm)