codex-imagegen

You cannot produce raster images. codex exec sessions can: the Codex CLI exposes a built-in image_gen.imagegen tool that generates and edits images using the user's existing ChatGPT authentication. No API key is required and no OPENAI_API_KEY is used — this runs on the plan the user already signed into with codex login.

Your job: turn the request into a precise image prompt, dispatch it, verify the file that comes back, and deliver it.

Step 0 — Preflight

1. command -v codex — if missing, stop and tell the user to install the Codex CLI (npm install -g @openai/codex) and run codex login.
2. Locate dispatch.sh as in codex-swarm Step 0 and use it for every job, including a single image — it resolves the current model for you.

Step 1 — Build the manifest

For each requested image, record as separate fields:

• purpose — what it is for (icon, hero, texture, OG card…)
• prompt — the detailed description you will send
• new or edit — edits need exact input paths
• destination — an absolute output path
• format — png, jpg, webp

Treat a user-specified path and format as immutable. Do not rename the file, change its extension, or pick a different directory without asking. If the user gave no path, choose a sensible project location (assets/, public/, docs/images/) and say which one you picked. If the exact path already exists and replacement was not requested, ask before overwriting.

Step 2 — Write the image prompt

Codex generates what you describe, so describe it fully. Cover: subject, composition and framing, style (flat vector, 3D render, photo, watercolor…), palette, lighting, background (state explicitly if it must be transparent), aspect ratio, and any text that must appear verbatim. Name what must not be there too — stray text, watermarks, borders.

For an edit, pass the exact absolute path of every input file and state what must stay unchanged.

Step 3 — Dispatch

One job per image. Images are independent, so a batch goes out in parallel.

Write one prompt file per image with the light tier header, then dispatch them together (a single image is a batch of one):

MODEL: light

Generate an image: <full prompt>. Save it to <ABS PATH>, then print a line
starting with TOUCHED: listing the file.

bash "$DISPATCH" --auto --timeout 20m 01-icon-light.prompt.md 02-icon-dark.prompt.md

Every prompt must end with: "Save it to <absolute path>, then print a line starting with TOUCHED: listing the file."

Rules:

• The sandbox must be workspace-write (--auto), and the destination must be inside the project — a read-only sandbox cannot save the file.
• One generation call per asset or variant. A retry is a new variant with its own path, never a second write to the same path.
• Two parallel jobs must never target the same output path.
• Use the light tier; the image model does the heavy lifting, not the reasoning model. The dispatcher resolves it to the current Codex model.
• Image generation takes ~40–90 seconds per asset. Allow at least 5 minutes per job and tell the user it is running.

Step 4 — Verify before you deliver

1. Confirm the file exists at the exact manifest path and is non-trivial in size (ls -l; a few hundred bytes means it failed).
2. View it with your image-reading tool. Check it against the requested subject, composition, text, and — for edits — that the invariants held.
3. Validate the real format from file contents, not the extension (file <path>).
4. If it is wrong, add one targeted variant to the manifest with a corrected prompt and its own path, dispatch that one job, and inspect again.

Report every final path, the validated format, and what you verified.

Blocking

If codex is missing, unauthenticated, or the job returns no file, stop. State which assets are incomplete and why. Do not describe an image you did not generate, do not claim a file exists without checking, and never ask the user to paste a secret into chat.